International Journal of Computer Science Issues

Transcription

1 IJCSI International Journal of Computer Science Issues Volume 7, Issue 6, November 2010 IJCSI PUBLICATION

2 IJCSI proceedings are currently indexed by: IJCSI PUBLICATION 2010

3 IJCSI Publicity Board 2010 Dr. Borislav D Dimitrov Department of General Practice, Royal College of Surgeons in Ireland Dublin, Ireland Dr. Vishal Goyal Department of Computer Science, Punjabi University Patiala, India Mr. Nehinbe Joshua University of Essex Colchester, Essex, UK Mr. Vassilis Papataxiarhis Department of Informatics and Telecommunications National and Kapodistrian University of Athens, Athens, Greece

4 EDITORIAL In this last edition of 2010, we bring forward issues from various dynamic computer science fields ranging from system performance, computer vision, artificial intelligence, software engineering, multimedia, pattern recognition, information retrieval, databases, security and networking among others. Considering the growing interest of academics worldwide to publish in IJCSI, we invite universities and institutions to partner with us to further encourage open-access publications. As always we thank all our reviewers for providing constructive comments on papers sent to them for review. This helps enormously in improving the quality of papers published in this issue. Google Scholar reported a large amount of cited papers published in IJCSI. We will continue to encourage the readers, authors and reviewers and the computer science scientific community and interested authors to continue citing papers published by the journal. Apart from availability of the full-texts from the journal website, all published papers are deposited in open-access repositories to make access easier and ensure continuous availability of its proceedings free of charge for all researchers. We are pleased to present IJCSI Volume 7, Issue 6, November 2010 (IJCSI Vol. 7, Issue 6). Out of the 196 paper submissions received, 54 papers were retained for publication. The acceptance rate for this issue is 27.55%. We wish you a Merry Christmas 2010 and a Happy New Year 2011! IJCSI Editorial Board November 2010 Issue IJCSI Publications

5 IJCSI Editorial Board 2010 Dr Tristan Vanrullen Chief Editor LPL, Laboratoire Parole et Langage - CNRS - Aix en Provence, France LABRI, Laboratoire Bordelais de Recherche en Informatique - INRIA - Bordeaux, France LEEE, Laboratoire d'esthétique et Expérimentations de l'espace - Université d'auvergne, France Dr Constantino Malagôn Associate Professor Nebrija University Spain Dr Lamia Fourati Chaari Associate Professor Multimedia and Informatics Higher Institute in SFAX Tunisia Dr Mokhtar Beldjehem Professor Sainte-Anne University Halifax, NS, Canada Dr Pascal Chatonnay Assistant Professor MaÎtre de Conférences Laboratoire d'informatique de l'université de Franche-Comté Université de Franche-Comté France Dr Karim Mohammed Rezaul Centre for Applied Internet Research (CAIR) Glyndwr University Wrexham, United Kingdom Dr Yee-Ming Chen Professor Department of Industrial Engineering and Management Yuan Ze University Taiwan

6 Dr Vishal Goyal Assistant Professor Department of Computer Science Punjabi University Patiala, India Dr Dalbir Singh Faculty of Information Science And Technology National University of Malaysia Malaysia Dr Natarajan Meghanathan Assistant Professor REU Program Director Department of Computer Science Jackson State University Jackson, USA Dr Deepak Laxmi Narasimha Department of Software Engineering, Faculty of Computer Science and Information Technology, University of Malaya, Kuala Lumpur, Malaysia Dr Navneet Agrawal Assistant Professor Department of ECE, College of Technology & Engineering, MPUAT, Udaipur Rajasthan, India Dr Shishir Kumar Department of Computer Science and Engineering, Jaypee University of Engineering & Technology Raghogarh, MP, India Dr P. K. Suri Professor Department of Computer Science & Applications, Kurukshetra University, Kurukshetra, India

7 Dr Paramjeet Singh Associate Professor GZS College of Engineering & Technology, India Dr Shaveta Rani Associate Professor GZS College of Engineering & Technology, India Dr G. Ganesan Professor Department of Mathematics, Adikavi Nannaya University, Rajahmundry, A.P, India Dr A. V. Senthil Kumar Department of MCA, Hindusthan College of Arts and Science, Coimbatore, Tamilnadu, India Dr T. V. Prasad Professor Department of Computer Science and Engineering, Lingaya's University Faridabad, Haryana, India Prof N. Jaisankar Assistant Professor School of Computing Sciences, VIT University Vellore, Tamilnadu, India

8 IJCSI Reviewers Committee 2010 Mr. Markus Schatten, University of Zagreb, Faculty of Organization and Informatics, Croatia Mr. Vassilis Papataxiarhis, Department of Informatics and Telecommunications, National and Kapodistrian University of Athens, Athens, Greece Dr Modestos Stavrakis, University of the Aegean, Greece Dr Fadi KHALIL, LAAS -- CNRS Laboratory, France Dr Dimitar Trajanov, Faculty of Electrical Engineering and Information technologies, ss. Cyril and Methodius Univesity - Skopje, Macedonia Dr Jinping Yuan, College of Information System and Management,National Univ. of Defense Tech., China Dr Alexis Lazanas, Ministry of Education, Greece Dr Stavroula Mougiakakou, University of Bern, ARTORG Center for Biomedical Engineering Research, Switzerland Dr Cyril de Runz, CReSTIC-SIC, IUT de Reims, University of Reims, France Mr. Pramodkumar P. Gupta, Dept of Bioinformatics, Dr D Y Patil University, India Dr Alireza Fereidunian, School of ECE, University of Tehran, Iran Mr. Fred Viezens, Otto-Von-Guericke-University Magdeburg, Germany Dr. Richard G. Bush, Lawrence Technological University, United States Dr. Ola Osunkoya, Information Security Architect, USA Mr. Kotsokostas N.Antonios, TEI Piraeus, Hellas Prof Steven Totosy de Zepetnek, U of Halle-Wittenberg & Purdue U & National Sun Yat-sen U, Germany, USA, Taiwan Mr. M Arif Siddiqui, Najran University, Saudi Arabia Ms. Ilknur Icke, The Graduate Center, City University of New York, USA Prof Miroslav Baca, Faculty of Organization and Informatics, University of Zagreb, Croatia Dr. Elvia Ruiz Beltrán, Instituto Tecnológico de Aguascalientes, Mexico Mr. Moustafa Banbouk, Engineer du Telecom, UAE Mr. Kevin P. Monaghan, Wayne State University, Detroit, Michigan, USA Ms. Moira Stephens, University of Sydney, Australia Ms. Maryam Feily, National Advanced IPv6 Centre of Excellence (NAV6), Universiti Sains Malaysia (USM), Malaysia Dr. Constantine YIALOURIS, Informatics Laboratory Agricultural University of Athens, Greece Mrs. Angeles Abella, U. de Montreal, Canada Dr. Patrizio Arrigo, CNR ISMAC, italy Mr. Anirban Mukhopadhyay, B.P.Poddar Institute of Management & Technology, India Mr. Dinesh Kumar, DAV Institute of Engineering & Technology, India Mr. Jorge L. Hernandez-Ardieta, INDRA SISTEMAS / University Carlos III of Madrid, Spain Mr. AliReza Shahrestani, University of Malaya (UM), National Advanced IPv6 Centre of Excellence (NAv6), Malaysia Mr. Blagoj Ristevski, Faculty of Administration and Information Systems Management - Bitola, Republic of Macedonia Mr. Mauricio Egidio Cantão, Department of Computer Science / University of São Paulo, Brazil Mr. Jules Ruis, Fractal Consultancy, The netherlands

9 Mr. Mohammad Iftekhar Husain, University at Buffalo, USA Dr. Deepak Laxmi Narasimha, Department of Software Engineering, Faculty of Computer Science and Information Technology, University of Malaya, Malaysia Dr. Paola Di Maio, DMEM University of Strathclyde, UK Dr. Bhanu Pratap Singh, Institute of Instrumentation Engineering, Kurukshetra University Kurukshetra, India Mr. Sana Ullah, Inha University, South Korea Mr. Cornelis Pieter Pieters, Condast, The Netherlands Dr. Amogh Kavimandan, The MathWorks Inc., USA Dr. Zhinan Zhou, Samsung Telecommunications America, USA Mr. Alberto de Santos Sierra, Universidad Politécnica de Madrid, Spain Dr. Md. Atiqur Rahman Ahad, Department of Applied Physics, Electronics & Communication Engineering (APECE), University of Dhaka, Bangladesh Dr. Charalampos Bratsas, Lab of Medical Informatics, Medical Faculty, Aristotle University, Thessaloniki, Greece Ms. Alexia Dini Kounoudes, Cyprus University of Technology, Cyprus Dr. Jorge A. Ruiz-Vanoye, Universidad Juárez Autónoma de Tabasco, Mexico Dr. Alejandro Fuentes Penna, Universidad Popular Autónoma del Estado de Puebla, México Dr. Ocotlán Díaz-Parra, Universidad Juárez Autónoma de Tabasco, México Mrs. Nantia Iakovidou, Aristotle University of Thessaloniki, Greece Mr. Vinay Chopra, DAV Institute of Engineering & Technology, Jalandhar Ms. Carmen Lastres, Universidad Politécnica de Madrid - Centre for Smart Environments, Spain Dr. Sanja Lazarova-Molnar, United Arab Emirates University, UAE Mr. Srikrishna Nudurumati, Imaging & Printing Group R&D Hub, Hewlett-Packard, India Dr. Olivier Nocent, CReSTIC/SIC, University of Reims, France Mr. Burak Cizmeci, Isik University, Turkey Dr. Carlos Jaime Barrios Hernandez, LIG (Laboratory Of Informatics of Grenoble), France Mr. Md. Rabiul Islam, Rajshahi university of Engineering & Technology (RUET), Bangladesh Dr. LAKHOUA Mohamed Najeh, ISSAT - Laboratory of Analysis and Control of Systems, Tunisia Dr. Alessandro Lavacchi, Department of Chemistry - University of Firenze, Italy Mr. Mungwe, University of Oldenburg, Germany Mr. Somnath Tagore, Dr D Y Patil University, India Ms. Xueqin Wang, ATCS, USA Dr. Borislav D Dimitrov, Department of General Practice, Royal College of Surgeons in Ireland, Dublin, Ireland Dr. Fondjo Fotou Franklin, Langston University, USA Dr. Vishal Goyal, Department of Computer Science, Punjabi University, Patiala, India Mr. Thomas J. Clancy, ACM, United States Dr. Ahmed Nabih Zaki Rashed, Dr. in Electronic Engineering, Faculty of Electronic Engineering, menouf 32951, Electronics and Electrical Communication Engineering Department, Menoufia university, EGYPT, EGYPT Dr. Rushed Kanawati, LIPN, France Mr. Koteshwar Rao, K G Reddy College Of ENGG.&TECH,CHILKUR, RR DIST.,AP, India Mr. M. Nagesh Kumar, Department of Electronics and Communication, J.S.S. research foundation, Mysore University, Mysore-6, India

10 Dr. Ibrahim Noha, Grenoble Informatics Laboratory, France Mr. Muhammad Yasir Qadri, University of Essex, UK Mr. Annadurai.P, KMCPGS, Lawspet, Pondicherry, India, (Aff. Pondicherry Univeristy, India Mr. E Munivel, CEDTI (Govt. of India), India Dr. Chitra Ganesh Desai, University of Pune, India Mr. Syed, Analytical Services & Materials, Inc., USA Mrs. Payal N. Raj, Veer South Gujarat University, India Mrs. Priti Maheshwary, Maulana Azad National Institute of Technology, Bhopal, India Mr. Mahesh Goyani, S.P. University, India, India Mr. Vinay Verma, Defence Avionics Research Establishment, DRDO, India Dr. George A. Papakostas, Democritus University of Thrace, Greece Mr. Abhijit Sanjiv Kulkarni, DARE, DRDO, India Mr. Kavi Kumar Khedo, University of Mauritius, Mauritius Dr. B. Sivaselvan, Indian Institute of Information Technology, Design & Manufacturing, Kancheepuram, IIT Madras Campus, India Dr. Partha Pratim Bhattacharya, Greater Kolkata College of Engineering and Management, West Bengal University of Technology, India Mr. Manish Maheshwari, Makhanlal C University of Journalism & Communication, India Dr. Siddhartha Kumar Khaitan, Iowa State University, USA Dr. Mandhapati Raju, General Motors Inc, USA Dr. M.Iqbal Saripan, Universiti Putra Malaysia, Malaysia Mr. Ahmad Shukri Mohd Noor, University Malaysia Terengganu, Malaysia Mr. Selvakuberan K, TATA Consultancy Services, India Dr. Smita Rajpal, Institute of Technology and Management, Gurgaon, India Mr. Rakesh Kachroo, Tata Consultancy Services, India Mr. Raman Kumar, National Institute of Technology, Jalandhar, Punjab., India Mr. Nitesh Sureja, S.P.University, India Dr. M. Emre Celebi, Louisiana State University, Shreveport, USA Dr. Aung Kyaw Oo, Defence Services Academy, Myanmar Mr. Sanjay P. Patel, Sankalchand Patel College of Engineering, Visnagar, Gujarat, India Dr. Pascal Fallavollita, Queens University, Canada Mr. Jitendra Agrawal, Rajiv Gandhi Technological University, Bhopal, MP, India Mr. Ismael Rafael Ponce Medellín, Cenidet (Centro Nacional de Investigación y Desarrollo Tecnológico), Mexico Mr. Supheakmungkol SARIN, Waseda University, Japan Mr. Shoukat Ullah, Govt. Post Graduate College Bannu, Pakistan Dr. Vivian Augustine, Telecom Zimbabwe, Zimbabwe Mrs. Mutalli Vatila, Offshore Business Philipines, Philipines Mr. Pankaj Kumar, SAMA, India Dr. Himanshu Aggarwal, Punjabi University,Patiala, India Dr. Vauvert Guillaume, Europages, France Prof Yee Ming Chen, Department of Industrial Engineering and Management, Yuan Ze University, Taiwan Dr. Constantino Malagón, Nebrija University, Spain Prof Kanwalvir Singh Dhindsa, B.B.S.B.Engg.College, Fatehgarh Sahib (Punjab), India

11 Mr. Angkoon Phinyomark, Prince of Singkla University, Thailand Ms. Nital H. Mistry, Veer Narmad South Gujarat University, Surat, India Dr. M.R.Sumalatha, Anna University, India Mr. Somesh Kumar Dewangan, Disha Institute of Management and Technology, India Mr. Raman Maini, Punjabi University, Patiala(Punjab) , India Dr. Abdelkader Outtagarts, Alcatel-Lucent Bell-Labs, France Prof Dr. Abdul Wahid, AKG Engg. College, Ghaziabad, India Mr. Prabu Mohandas, Anna University/Adhiyamaan College of Engineering, india Dr. Manish Kumar Jindal, Panjab University Regional Centre, Muktsar, India Prof Mydhili K Nair, M S Ramaiah Institute of Technnology, Bangalore, India Dr. C. Suresh Gnana Dhas, VelTech MultiTech Dr.Rangarajan Dr.Sagunthala Engineering College,Chennai,Tamilnadu, India Prof Akash Rajak, Krishna Institute of Engineering and Technology, Ghaziabad, India Mr. Ajay Kumar Shrivastava, Krishna Institute of Engineering & Technology, Ghaziabad, India Mr. Deo Prakash, SMVD University, Kakryal(J&K), India Dr. Vu Thanh Nguyen, University of Information Technology HoChiMinh City, VietNam Prof Deo Prakash, SMVD University (A Technical University open on I.I.T. Pattern) Kakryal (J&K), India Dr. Navneet Agrawal, Dept. of ECE, College of Technology & Engineering, MPUAT, Udaipur Rajasthan, India Mr. Sufal Das, Sikkim Manipal Institute of Technology, India Mr. Anil Kumar, Sikkim Manipal Institute of Technology, India Dr. B. Prasanalakshmi, King Saud University, Saudi Arabia. Dr. K D Verma, S.V. (P.G.) College, Aligarh, India Mr. Mohd Nazri Ismail, System and Networking Department, University of Kuala Lumpur (UniKL), Malaysia Dr. Nguyen Tuan Dang, University of Information Technology, Vietnam National University Ho Chi Minh city, Vietnam Dr. Abdul Aziz, University of Central Punjab, Pakistan Dr. P. Vasudeva Reddy, Andhra University, India Mrs. Savvas A. Chatzichristofis, Democritus University of Thrace, Greece Mr. Marcio Dorn, Federal University of Rio Grande do Sul - UFRGS Institute of Informatics, Brazil Mr. Luca Mazzola, University of Lugano, Switzerland Mr. Nadeem Mahmood, Department of Computer Science, University of Karachi, Pakistan Mr. Hafeez Ullah Amin, Kohat University of Science & Technology, Pakistan Dr. Professor Vikram Singh, Ch. Devi Lal University, Sirsa (Haryana), India Mr. M. Azath, Calicut/Mets School of Enginerring, India Dr. J. Hanumanthappa, DoS in CS, University of Mysore, India Dr. Shahanawaj Ahamad, Department of Computer Science, King Saud University, Saudi Arabia Dr. K. Duraiswamy, K. S. Rangasamy College of Technology, India Prof. Dr Mazlina Esa, Universiti Teknologi Malaysia, Malaysia Dr. P. Vasant, Power Control Optimization (Global), Malaysia Dr. Taner Tuncer, Firat University, Turkey Dr. Norrozila Sulaiman, University Malaysia Pahang, Malaysia Prof. S K Gupta, BCET, Guradspur, India

12 Dr. Latha Parameswaran, Amrita Vishwa Vidyapeetham, India Mr. M. Azath, Anna University, India Dr. P. Suresh Varma, Adikavi Nannaya University, India Prof. V. N. Kamalesh, JSS Academy of Technical Education, India Dr. D Gunaseelan, Ibri College of Technology, Oman Mr. Sanjay Kumar Anand, CDAC, India Mr. Akshat Verma, CDAC, India Mrs. Fazeela Tunnisa, Najran University, Kingdom of Saudi Arabia Mr. Hasan Asil, Islamic Azad University Tabriz Branch (Azarshahr), Iran Prof. Dr Sajal Kabiraj, Fr. C Rodrigues Institute of Management Studies (Affiliated to University of Mumbai, India), India Mr. Syed Fawad Mustafa, GAC Center, Shandong University, China Dr. Natarajan Meghanathan, Jackson State University, Jackson, MS, USA Prof. Selvakani Kandeeban, Francis Xavier Engineering College, India Mr. Tohid Sedghi, Urmia University, Iran Dr. S. Sasikumar, PSNA College of Engg and Tech, Dindigul, India Dr. Anupam Shukla, Indian Institute of Information Technology and Management Gwalior, India Mr. Rahul Kala, Indian Institute of Inforamtion Technology and Management Gwalior, India Dr. A V Nikolov, National University of Lesotho, Lesotho Mr. Kamal Sarkar, Department of Computer Science and Engineering, Jadavpur University, India Dr. Mokhled S. AlTarawneh, Computer Engineering Dept., Faculty of Engineering, Mutah University, Jordan, Jordan Prof. Sattar J Aboud, Iraqi Council of Representatives, Iraq-Baghdad Dr. Prasant Kumar Pattnaik, Department of CSE, KIST, India Dr. Mohammed Amoon, King Saud University, Saudi Arabia Dr. Tsvetanka Georgieva, Department of Information Technologies, St. Cyril and St. Methodius University of Veliko Tarnovo, Bulgaria Dr. Eva Volna, University of Ostrava, Czech Republic Mr. Ujjal Marjit, University of Kalyani, West-Bengal, India Dr. Prasant Kumar Pattnaik, KIST,Bhubaneswar,India, India Dr. Guezouri Mustapha, Department of Electronics, Faculty of Electrical Engineering, University of Science and Technology (USTO), Oran, Algeria Mr. Maniyar Shiraz Ahmed, Najran University, Najran, Saudi Arabia Dr. Sreedhar Reddy, JNTU, SSIETW, Hyderabad, India Mr. Bala Dhandayuthapani Veerasamy, Mekelle University, Ethiopa Mr. Arash Habibi Lashkari, University of Malaya (UM), Malaysia Mr. Rajesh Prasad, LDC Institute of Technical Studies, Allahabad, India Ms. Habib Izadkhah, Tabriz University, Iran Dr. Lokesh Kumar Sharma, Chhattisgarh Swami Vivekanand Technical University Bhilai, India Mr. Kuldeep Yadav, IIIT Delhi, India Dr. Naoufel Kraiem, Institut Superieur d'informatique, Tunisia Prof. Frank Ortmeier, Otto-von-Guericke-Universitaet Magdeburg, Germany Mr. Ashraf Aljammal, USM, Malaysia Mrs. Amandeep Kaur, Department of Computer Science, Punjabi University, Patiala, Punjab, India Mr. Babak Basharirad, University Technology of Malaysia, Malaysia

13 Mr. Avinash singh, Kiet Ghaziabad, India Dr. Miguel Vargas-Lombardo, Technological University of Panama, Panama Dr. Tuncay Sevindik, Firat University, Turkey Ms. Pavai Kandavelu, Anna University Chennai, India Mr. Ravish Khichar, Global Institute of Technology, India Mr Aos Alaa Zaidan Ansaef, Multimedia University, Cyberjaya, Malaysia Dr. Awadhesh Kumar Sharma, Dept. of CSE, MMM Engg College, Gorakhpur , UP, India Mr. Qasim Siddique, FUIEMS, Pakistan Dr. Le Hoang Thai, University of Science, Vietnam National University - Ho Chi Minh City, Vietnam Dr. Saravanan C, NIT, Durgapur, India Dr. Vijay Kumar Mago, DAV College, Jalandhar, India Dr. Do Van Nhon, University of Information Technology, Vietnam Mr. Georgios Kioumourtzis, University of Patras, Greece Mr. Amol D.Potgantwar, SITRC Nasik, India Mr. Lesedi Melton Masisi, Council for Scientific and Industrial Research, South Africa Dr. Karthik.S, Department of Computer Science & Engineering, SNS College of Technology, India Mr. Nafiz Imtiaz Bin Hamid, Department of Electrical and Electronic Engineering, Islamic University of Technology (IUT), Bangladesh Mr. Muhammad Imran Khan, Universiti Teknologi PETRONAS, Malaysia Dr. Abdul Kareem M. Radhi, Information Engineering - Nahrin University, Iraq Dr. Mohd Nazri Ismail, University of Kuala Lumpur, Malaysia Dr. Manuj Darbari, BBDNITM, Institute of Technology, A-649, Indira Nagar, Lucknow , India Ms. Izerrouken, INP-IRIT, France Mr. Nitin Ashokrao Naik, Dept. of Computer Science, Yeshwant Mahavidyalaya, Nanded, India Mr. Nikhil Raj, National Institute of Technology, Kurukshetra, India Prof. Maher Ben Jemaa, National School of Engineers of Sfax, Tunisia Prof. Rajeshwar Singh, BRCM College of Engineering and Technology, Bahal Bhiwani, Haryana, India Mr. Gaurav Kumar, Department of Computer Applications, Chitkara Institute of Engineering and Technology, Rajpura, Punjab, India Mr. Ajeet Kumar Pandey, Indian Institute of Technology, Kharagpur, India Mr. Rajiv Phougat, IBM Corporation, USA Mrs. Aysha V, College of Applied Science Pattuvam affiliated with Kannur University, India Dr. Debotosh Bhattacharjee, Department of Computer Science and Engineering, Jadavpur University, Kolkata , India Dr. Neelam Srivastava, Institute of engineering & Technology, Lucknow, India Prof. Sweta Verma, Galgotia's College of Engineering & Technology, Greater Noida, India Mr. Harminder Singh BIndra, MIMIT, INDIA Dr. Lokesh Kumar Sharma, Chhattisgarh Swami Vivekanand Technical University, Bhilai, India Mr. Tarun Kumar, U.P. Technical University/Radha Govinend Engg. College, India Mr. Tirthraj Rai, Jawahar Lal Nehru University, New Delhi, India Mr. Akhilesh Tiwari, Madhav Institute of Technology & Science, India Mr. Dakshina Ranjan Kisku, Dr. B. C. Roy Engineering College, WBUT, India Ms. Anu Suneja, Maharshi Markandeshwar University, Mullana, Haryana, India Mr. Munish Kumar Jindal, Punjabi University Regional Centre, Jaito (Faridkot), India

14 Dr. Ashraf Bany Mohammed, Management Information Systems Department, Faculty of Administrative and Financial Sciences, Petra University, Jordan Mrs. Jyoti Jain, R.G.P.V. Bhopal, India Dr. Lamia Chaari, SFAX University, Tunisia Mr. Akhter Raza Syed, Department of Computer Science, University of Karachi, Pakistan Prof. Khubaib Ahmed Qureshi, Information Technology Department, HIMS, Hamdard University, Pakistan Prof. Boubker Sbihi, Ecole des Sciences de L'Information, Morocco Dr. S. M. Riazul Islam, Inha University, South Korea Prof. Lokhande S.N., S.R.T.M.University, Nanded (MH), India Dr. Vijay H Mankar, Dept. of Electronics, Govt. Polytechnic, Nagpur, India Dr. M. Sreedhar Reddy, JNTU, Hyderabad, SSIETW, India Mr. Ojesanmi Olusegun, Ajayi Crowther University, Oyo, Nigeria Ms. Mamta Juneja, RBIEBT, PTU, India Dr. Ekta Walia Bhullar, Maharishi Markandeshwar University, Mullana Ambala (Haryana), India Prof. Chandra Mohan, John Bosco Engineering College, India Mr. Nitin A. Naik, Yeshwant Mahavidyalaya, Nanded, India Mr. Sunil Kashibarao Nayak, Bahirji Smarak Mahavidyalaya, Basmathnagar Dist-Hingoli., India Prof. Rakesh.L, Vijetha Institute of Technology, Bangalore, India Mr B. M. Patil, Indian Institute of Technology, Roorkee, Uttarakhand, India Mr. Thipendra Pal Singh, Sharda University, K.P. III, Greater Noida, Uttar Pradesh, India Prof. Chandra Mohan, John Bosco Engg College, India Mr. Hadi Saboohi, University of Malaya - Faculty of Computer Science and Information Technology, Malaysia Dr. R. Baskaran, Anna University, India Dr. Wichian Sittiprapaporn, Mahasarakham University College of Music, Thailand Mr. Lai Khin Wee, Universiti Teknologi Malaysia, Malaysia Dr. Kamaljit I. Lakhtaria, Atmiya Institute of Technology, India Mrs. Inderpreet Kaur, PTU, Jalandhar, India Mr. Iqbaldeep Kaur, PTU / RBIEBT, India Mrs. Vasudha Bahl, Maharaja Agrasen Institute of Technology, Delhi, India Prof. Vinay Uttamrao Kale, P.R.M. Institute of Technology & Research, Badnera, Amravati, Maharashtra, India Mr. Suhas J Manangi, Microsoft, India Ms. Anna Kuzio, Adam Mickiewicz University, School of English, Poland Mr. Vikas Singla, Malout Institute of Management & Information Technology, Malout, Punjab, India, India Dr. Dalbir Singh, Faculty of Information Science And Technology, National University of Malaysia, Malaysia Dr. Saurabh Mukherjee, PIM, Jiwaji University, Gwalior, M.P, India Dr. Debojyoti Mitra, Sir Padampat Singhania University, India Prof. Rachit Garg, Department of Computer Science, L K College, India Dr. Arun Kumar Gupta, M.S. College, Saharanpur, India Dr. Todor Todorov, Institute of Mathematics and Informatics, Bulgarian Academy of Sciences, Bulgaria

15 Mr. Akhter Raza Syed, University of Karachi, Pakistan Mrs. Manjula K A, Kannur University, India Prof. M. Saleem Babu, Department of Computer Science and Engineering, Vel Tech University, Chennai, India Dr. Rajesh Kumar Tiwari, GLA Institute of Technology, India Dr. V. Nagarajan, SMVEC, Pondicherry university, India Mr. Rakesh Kumar, Indian Institute of Technology Roorkee, India Prof. Amit Verma, PTU/RBIEBT, India Mr. Sohan Purohit, University of Massachusetts Lowell, USA Mr. Anand Kumar, AMC Engineering College, Bangalore, India Dr. Samir Abdelrahman, Computer Science Department, Cairo University, Egypt Dr. Rama Prasad V Vaddella, Sree Vidyanikethan Engineering College, India Prof. Jyoti Prakash Singh, Academy of Technology, India Mr. Peyman Taher, Oklahoma State University, USA Dr. S Srinivasan, PDM College of Engineering, India Mr. Muhammad Zakarya, CIIT, Pakistan Mr. Williamjeet Singh, Chitkara Institute of Engineering and Technology, India Mr. G.Jeyakumar, Amrita School of Engineering, India Mr. Harmunish Taneja, Maharishi Markandeshwar University, Mullana, Ambala, Haryana, India Dr. Sin-Ban Ho, Faculty of IT, Multimedia University, Malaysia Mrs. Doreen Hephzibah Miriam, Anna University, Chennai, India Mrs. Mitu Dhull, GNKITMS Yamuna Nagar Haryana, India Mr. Neetesh Gupta, Technocrats Inst. of Technology, Bhopal, India Ms. A. Lavanya, Manipal University, Karnataka, India Ms. D. Pravallika, Manipal University, Karnataka, India Prof. Ashutosh Kumar Dubey, Assistant Professor, India Mr. Ranjit Singh, Apeejay Institute of Management, Jalandhar, India Mr. Prasad S.Halgaonkar, MIT, Pune University, India Mr. Anand Sharma, MITS, Lakshmangarh, Sikar (Rajasthan), India Mr. Amit Kumar, Jaypee University of Engineering and Technology, India Prof. Vasavi Bande, Computer Science and Engneering, Hyderabad Institute of Technology and Management, India Dr. Jagdish Lal Raheja, Central Electronics Engineering Research Institute, India Mr G. Appasami, Dept. of CSE, Dr. Pauls Engineering College, Anna University - Chennai, India Mr Vimal Mishra, U.P. Technical Education, Allahabad, India Dr. Arti Arya, PES School of Engineering, Bangalore (under VTU, Belgaum, Karnataka), India Mr. Pawan Jindal, J.U.E.T. Guna, M.P., India Dr. P. K. Suri, Department of Computer Science & Applications, Kurukshetra University, Kurukshetra, India Dr. Syed Akhter Hossain, Daffodil International University, Bangladesh Mr. Nasim Qaisar, Federal Urdu Univetrsity of Arts, Science and Technology, Pakistan Mr. Mohit Jain, Maharaja Surajmal Institute of Technology (Affiliated to Guru Gobind Singh Indraprastha University, New Delhi), India Dr. Shaveta Rani, GZS College of Engineering & Technology, India Dr. Paramjeet Singh, GZS College of Engineering & Technology, India

16 Prof. T Venkat Narayana Rao, Department of CSE, Hyderabad Institute of Technology and Management, India Mr. Vikas Gupta, CDLM Government Engineering College, Panniwala Mota, India Dr Juan José Martínez Castillo, University of Yacambu, Venezuela Mr Kunwar S. Vaisla, Department of Computer Science & Engineering, BCT Kumaon Engineering College, India Prof. Manpreet Singh, M. M. Engg. College, M. M. University, Haryana, India Mr. Syed Imran, University College Cork, Ireland Dr. Namfon Assawamekin, University of the Thai Chamber of Commerce, Thailand Dr. Shahaboddin Shamshirband, Islamic Azad University, Iran Dr. Mohamed Ali Mahjoub, University of Monastir, Tunisia Mr. Adis Medic, Infosys ltd, Bosnia and Herzegovina Mr Swarup Roy, Department of Information Technology, North Eastern Hill University, Umshing, Shillong , Meghalaya, India

17 TABLE OF CONTENTS 1. Mobile agent driven by aspect Youssef Hannad and Fabrice Mourlin 2. A Modified Algorithm of Bare Bones Particle Swarm Optimization Horng-I Hsieh and Tian-Shyug Lee 3. Web Personalization of Indian e-commerce Websites using Classification Methodologies Agarwal Devendera, Tripathi S.P and Singh J.B. 4. High Accuracy Myanmar Handwritten Character Recognition using Hybrid approach through MICR and Neural Network Yadana Thein and San Su Su Yee 5. Ontology Based Agent Communication in Resource Allocation and Monitoring Manish Arora and M. Syamala Devi 6. A Novel DSS Framework for E-government A.M. Riad, Hazem M. El-Bakry and Gamal H. El-Adl 7. Real-time Error Measurement System for MVB Protocol Su Goog Shon and Soo Mi Yang 8. General Database Infrastructure for Image Retrieval Carlos Alvez and Aldo Vecchietti 9. New approach using Bayesian Network to improve content based image classification systems Khlifia Jayech and Mohamed Ali Mahjoub 10. SbSAD: An Integrated Service-based Software Design Framework Mohamed Dbouk, Hamid Mcheick and Ihab Sbeity 11. Success Rules of OSS Projects using Datamining 3-Itemset Association Rule Andi Wahju Rahardjo Emanuel, Retantyo Wardoyo, Jazi Eko Istiyanto and Khabib Mustofa 12. Scalable Contents Delivery System with Dynamic Server Deployment Yuko Kamiya, Toshihiko Shimokawa, Fuminori Tanizaki and Norihiko Yoshida 13. A Survey on Performance Evaluation of Object Detection Techniques in Digital Image Processing J. Komala Lakshmi and M. Punithavalli Pg 1-11 Pg Pg Pg Pg Pg Pg Pg Pg Pg Pg Pg Pg 86-94

18 14. Optimal Provisioning of Resource in a Cloud Service Yee Ming Chen and Shin-Ying Tsai 15. Presenting a New Routing Protocol for increasing Lifetime of Sensor Network Vahid Majid Nezhad and Bager Zarei 16. A Knowledge Management Model to Improve Information Security Yogesh Kumar Mittal, Santanu Roy and Manu Saxena 17. TH*: Scalable Distributed Trie Hashing Aridj Mohamed and Zegour Djamel Edinne 18. Attacks in WEB Based Embedded Applications C. Yaashuwanth and R. Ramesh 19. A Tool for Qualitative Causal Reasoning On Complex Systems Tahar Guerram, Ramdane Maamri and Zaidi Sahnoun 20. Health Smart Home Ahmad Choukeir, Batoul Fneish, Nour Zaarour, Walid Fahs and Mohammad Ayache 21. SD-AODV: A Protocol for Secure and Dynamic Data Dissemination in Mobile Ad Hoc Network Rajender Nath and Pankaj Kumar Sehgal 22. A Radio Based Intelligent Railway Grade Crossing System to Avoid Collision Sheikh Shanawaz Mostafa, Md. Mahbub Hossian, Khondker Jahid Reza and Gazi Maniur Rashid 23. Effective Approaches For Extraction Of Keywords Jasmeen Kaur and Vishal Gupta 24. Image Splicing Detection Using Inherent Lens Radial Distortion H. R. Chennamma and Lalitha Rangarajan 25. Designing A Re-Configurable Fractional Fourier Transform Architecture Using Systolic Array Anal Acharya and Soumen Mukherjee 26. Syllables Selection for the Development of Speech Database for Punjabi TTS System Parminder Singh and Gurpreet Singh Lehal 27. Personalized access towards a mobile neuroscience Youssouf El Allioui and Omar El Beqqali 28. High Performance Direct Torque Control of Induction Motor Drives Using Space Vector Modulation S. Allirani and V. Jagannathan Pg Pg Pg Pg Pg Pg Pg Pg Pg Pg Pg Pg Pg Pg Pg

19 29. A new Morphological Approach for Noise Removal cum Edge Detection M. Rama Bai, V. Venkata Krishna and J. SreeDevi 30. Modeling ODP Policies by using event-b Belhaj Hafid, Mohamed Bouhdadi and Said Elhajji 31. A New Security Paradigm of File Sharing Seifedine Kadry 32. An Enhanced Algorithm Of The Statistical Training Method In Boosting-Based Face Detection Said Belkouch, Mounir Bahtat, Abdellah Ait Ouahman and M. M'rabet Hassani 33. Personnel Audit Using a Forensic Mining Technique Adesesan B. Adeyemo and Oluwafemi Oriola 34. A RSS Based Adaptive Hand-Off Management Scheme In Heterogeneous Networks Debabrata Sarddar, Shovan Maity, Arnab Raha, Ramesh Jana, Utpal Biswas and Mrinal Kanti Naskar 35. A survey of Named Entity Recognition in English and other Indian Languages Darvinder Kaur and Vishal Gupta 36. A Framework for Prefetching Relevant Web Pages using Predictive Prefetching Engine (PPE) Jyoti, Ashok Kale Sharma and Amit Goel 37. Genetic Algorithms as Virtual User Managers Rasmiprava Singh, Snehalata Baede and Sujata Khobragade 38. Performance Analysis of Peak-to-Average Power Ratio Reduction Techniques for Wireless Communication Using OFDM Signals Pawan Sharma and Seema Verma 39. To Design Voice Control Keyboard System using Speech Application Programming Interface Md. Sipon Mia and Tapan Kumar Godder 40. Combining of Spatial and Frequency Domain Transformation With The Effect of Using and Non-Using Adaptive Quantization for Image Compression Alan Anwer Abdulla 41. Automatic Computation for Pressure Controlled Intermittent Coronary Sinus Occlusion Loay Alzubaidi, Werner Mohl and Frank Rattay 42. Computer Science in Education Irshad Ullah Pg Pg Pg Pg Pg Pg Pg Pg Pg Pg Pg Pg Pg Pg

20 43. WLAN Security Flaw: Cracking 64 bit WEP Key Anil Kumar Singh, Bharat Mishra and Sandeep Singh 44. Digitizing the Forest Resource Map Using ArcGIS K. R. Manjula, S. Jyothi and S. Anand Kumar Varma 45. Digital Watermarking in Discrete Wavelet Transformation - Survey Rama Seshagiri Rao C., Sampath Kumar M. and Prakasam T. 46. A Methodology for Aiding Investment Decision between Assets in Stock Markets Using Artificial Neural Network P. N. Kumar, Rahul Seshadri G., A. Hariharan, V. P. Mohandas and P. Balasubramanian 47. Use of XML to Persist and Transfer Offline Data For Personalized Profiles Sandeep Chawla and Suchita Goyal 48. Role of Knowledge Management in Enhancing Information Security Yogesh Kumar Mittal, Santanu Roy and Manu Saxena 49. Cloud Computing for Managing Apparel and Garment Supply Chains - an Empirical study of Implementation Frame Work A.K. Damodaram and K. Ravindranath 50. Retrieval of average sum of plans and degree coefficient between genes in distributed Query Processing Sambit Kumar Mishra and Srikanta Pattnaik 51. A Goal Based Approach for QFD Refinement in Systematizing and Identifying Business Process Requirements Roger Atsa Etoundi, Marcel Fouda Ndjodo and Atouba Christian Lopez 52. Application of some Retrieved Information Method on Internet Vu Thanh Nguyen and Nguyen Quoc Thinh 53. Mixed-Myanmar and English Character Recognition with Formatting Yadana Thein and Cherry Maung 54. Personalized Online Learning with Ontological Approach Bernard Renaldy Suteja, Suryo Guritno, Retantyo Wardoyo and Ahmad Ashari Pg Pg Pg Pg Pg Pg Pg Pg Pg Pg Pg Pg

21 1 Mobile agent driven by aspect Youssef Hannad 1, Fabrice Mourlin 2 1 LACL, Laboratory Algorithm Complexity Logics Computer Science department, Paris 12 University, Creteil, 94010, France 2 LACL, Laboratory Algorithm Complexity Logics Computer Science department, Paris 12 University, Creteil, 94010, France Abstract Domain application of mobile agents is quite large. They are used for network management and the monitoring of complex architecture. Mobile agent is also essential into specific software architecture such that adaptable grid architecture. Even if the concept of mobile agent seems to be obvious, the development is always complex because it needs to understand network features but also security features and negotiation algorithms. We present a work about an application of aspects dedicated to mobile agent development over a local network. At this level, the underlying protocol is called jini and allows managing several essential concepts such that short transaction and permission management. Three subsets of aspects are defined in this work. A part is for the description of agent host and its security level, accessible resource, etc. A second part is about mobile agent and their collaboration. This means how they can operate on an agent host with the respect of the execution context. All the results are illustrated through a distributed monitoring application called DMA. Its main objective is the observation of component servers. Keywords: mobile agent, aspect programming, distributed application. 1. Introduction In a distributed context, where software shares resources, also a key concept is adaptability. Behind this word, several problems are hidden. Such as examples, numerical application need to access to computing resources over a network. These resources can be locked by another application, or the rights of the numerical application are not sufficient for the operation. In that case, this local anomaly can involve a global perturbation. It is essential to solve this problem locally. A central approach will involve a lot of message exchanges, every pieces of the distributed application will be touched by a local resource access violation. Therefore, a solution should be found locally. It means that a strategy has to be deployed for finding another resource for instance, or for acquiring new access permissions. For network monitoring, similar problems occur. When an administrator wants to observe the hosts and the state of the applications which are deployed, a part of an application can be inaccessible, also a remote observation can not be computed. Idem, administrator can not reboot the whole application because of a local anomaly. A diagnostic can be found and a solution can be set. For instance, this solution could be to move the local activity to another host or to save the current state of the activity, then to restart local processes. Similar as before, the decision graph has to be efficient if we want to solve a problem without to many perturbations. Previous examples highlight several features. First, adaptable behavior means that diagnostic and action have to be decided locally to the location where the problem happens. Secondly, action has to be effective: if an activity is not realized, it should be realized somewhere else. Often, problems are due to a service failure or a breakdown of material. Also, migration is a solution to replay an activity into another context. When a set of actions is moved from one node of the network to another, a collection of properties have to be checked. We can divide these properties into three subsets. A first one is about network characteristics, this means material description, protocol configuration and details on message routes. A second subset describes security permissions; this is essential for the negotiation step. When an activity moves from a node to another one, this activity has to be accepted by a host node. Because each node has its own permission strategy, the host negotiates with the activity to know whether or not this activity can be imported. Of course, this depends on what the activity wish to do and the resource it needs to use. Last subset of properties is about administration of the migration. When

22 2 an activity changes its locality, this can impact other activities, especially for message exchanges and more generally for monitoring. We can easily sum up that implementation of such mobile code is a complex task and several technologies can be applied for that objective [1]. This document is divided into five parts with first an introduction about mobile agent programming and application. Secondly, we detail our choice about aspect programming and the reasons for their introduction. Thirdly, we present our design strategy through UML notation and then we present our monitoring application. By the end, we present results of data collection and also the expressive power of message format 1.1 Mobile Agent Programming Our past experience on mobile agent programming allows us to compare frameworks and also to choose technology depending on its limits. For instance, JADE [2] needs to install services on host node; users cannot decide whether a service is mandatory or not. When runtime environment is restricted about CPU capacity of deployment ability, too many services involves side effect. Tryllian Mobile Agent Technology provides the necessary elements such as security [3], mobility, decision for activity management over a network, but administration is missing. Also, it becomes difficult to diagnostic precisely blocking into an agent application. The framework SOMA [4] has been designed to achieve two main objectives: security and interoperability. Application based on SOMA, can interoperate with different application components designed with different programming styles; it grants interoperability by closely considering compliance with CORBA and MASIF. But this brings a large amount of development rules based on design patterns. Also, software development costs much more with SOMA than other framework like JADE or JAVact [5]. Developers need to use tools for checking if design patterns are correctly applied, but such tools do not exist, also each developer has to assimilate a lot of knowledge to have programming level. Our main observation is this lack of help; it appears that a standard approach of agent mobile programming has to be defined to reduce cost of development. We present in the current document our contribution to the domain of mobile agent development based on the use of aspects approach. Our architecture constraints, described previously, are based on these three subsets and their controls imply particular properties of aspect programming. 1.2 Application domain of mobile agent system Mobile agents are software abstractions that can migrate across the network. This property has a large spectrum of applications. As mentioned previously, network administration was the first domain where mobile agents are considered as a probe or a spy which provides details about activities of a remote computer or device. A first example is an agent which gauges the load on specific ports. When a threshold is achieved, then agent can export a messenger (or mobile agent) to server with data about the alert. Mobile agents are also used into intrusion detection system. They have two roles. At the beginning, mobile agents are deployed from an agent server onto hosts where controls have to be done. For instance, an agent can observe protocol login and filters users which try to connect too many times. Regularly, agents notify server to ensure that they are always alive. This is essential when an attack occurs, because the invasion starts by killing agent which observes the protocol. When an agent is not alive, the server exports immediately another one. Thereby the safety service continues until the next attack. We have developed project based on AAFID (Autonomous Agents for Intrusion Detection) [6] strategy, this means a hierarchical architecture based on four kind of agent: monitor, decider, guardian, and filter. Some of them are mobile: decider and guardian, the others are static. Our prototype was used to observe activities on network local to teaching department. The results were surprising about the attack number even on computer without any strategic resource. The robustness of our approach has been enhanced and the concept of mobile agent has been linked for the first time to adaptable context. But a prototype is not enough to affirm general assertion and other examples are developed over the past decade. To destroy an agent is similar to material failure. We encountered this problem with grid computing where a computation is distributed over the nodes of a grid. The input data are scattered on the nodes, but when a part of the whole computation has an exception, the global result is touched. This exception can come from material or software. To solve this problem of exception management, we defined a software architecture based on a set of mobile agents, called computing space [7]. A server manages not only all components of numerical code, but also all the input data. When a processor of the grid becomes free, a mobile agent is exported onto that node with corresponding input data. When this part of the computation is done, the state of this part is set into the space computing. When all the parts of computation are realized, the termination of the computing case is detected. We use this architecture for several case studies: Choleski computing [8], FDTD computing, Laplace resolution [9]. With these examples, we enhance the idea that mobile agent can adapt a code to its working context: exploitation of free computing resources, replay a piece of computation previously interrupted, etc. We highlight also a new facet of mobile agent called local negotiation. Before agent is exported on a node, a negotiation is established between

23 3 agent host and mobile agent to know whether or not the computation can take place on this specific node. Control can be about resource access, location constraints, or security permissions, etc. This new facet of mobile agent programming add a layer of concepts. This increases development complexity and a project can become hard to maintain if its authors do not apply clear development rules and code convention. Also people who have enough experience in software development can accept and understand easily these constraints but it could be more useful to strive to identify a set of best practices. Then, software engineering tools could help developers to apply them. In order to prepare such tools, we have to isolate concept families. 1.3 Concept partitioning Today, everyone knows aspect-oriented programming as a new approach to software design. But its advantages are not so used such as modularity, concept isolation, etc. Other approaches, including structured programming and object-oriented programming need user experience to obtain same results. Also we can consider aspect-oriented programming as complement to traditional approaches. It introduces the mechanism of cross cutting for expressing concerns like action migration and automatically incorporating the associated code into the whole system. Thus, it enhances our ability to express the separation of concerns necessary for a well-designed, maintainable mobile agent system. Some concerns are appropriately expressed as encapsulated agents, or components. As an example, we can note the behavior of a mobile agent onto a node. Others are best expressed as cross-cutting concerns, for example, the precise route of a mobile agent through the client network. A difficulty remains: the identification of places in the code where we want to insert the route description. This is called defining join points. Where the aspects are used much depend one what they are used for. For instance, route definition of mobile agent is key information which can be defined at the declaration step. Again, security constraints are good candidates for becoming an aspect definition. This is functionality that is often used in agent but not a part of the normal business logic. This is also aspects that can be reused in many mobile agents and also be reused in different applications. The main design improvement we get with aspect oriented programming is better modularization. Redundant code can be placed in an aspect instead of copied to all agent definitions that need it. Developer can concentrate on putting the business logic in the agent definitions and the other part can be handled with aspects. Maintenance of mobile agent application can be improved by an aspect approach programming. This makes the code easier to read and observe. This is particularly essential during debugging phase. A disadvantage is about the understanding of join points and aspect definition. The code can become harder to follow because this can specially be a problem if the design is changed later during the lifecycle of application and functionality can be added with aspects. In our case study, we can sum up with three groups of facets: mobility, security and administration. We consider this objective as a basis of our framework. Moreover, limits can be added in a first approach: dynamicity of migration, permission evolution or evolving observation can be considered as advanced concepts. Also, we are interested in their application but in a second development step. The first development step is about migration, negotiation and agent management. This involves technical choice about aspect definition tool. 2. Aspect definition tool Aspect definition raises technical constraint depending on kind of aspect. Most of aspect compilers are based on a clear principle: code injection. But the strategy to apply this principle can vary from one implementation to another. A large set of aspect weaver works on source code or byte code. This means that they modify project by injection of technical code into source or byte code. This is useful for generation of XML descriptor into a J2EE project for instance. In our context of mobile agent application, when the migrations of agent are predefined, technical code can be generated from agent definition. But because the effect of the weaver occurs before the execution of the project, the limit of this approach happens when the travel of mobile agent changes during its work. The weaver is not able to change what it was previously generated. Tools operate with that kind of mechanism: such that AspectJ (first version) [10], JAC [11] or Hyper/J [12]. To solve the lack of dynamicity, a tool like AspectWerkz [13] or JMangler [14] proposes to apply aspect weaver when the class is loading. In a distributed system, this approach is interesting because each agent host has its own class loader. Also, we obtain a new behavior where the effect of advices can be used not only once, but several times depending of the travel of mobile agent. It is dynamic in the sense that it is possible to add, remove and restructure advices as well as swapping the implementation of the introductions at runtime. Like before, the byte code is touched at the loading step but a new limit is achieved. Now, agent behavior is determined at its entrance on to an agent host. This means that its local activity and its next migration are fixed when it arrives on an agent host. In other words, this mechanism prohibits any change into local agent behavior and also with the decision to navigate over the network.

24 4 2.1 Dynamic weaving process This approach is more powerful, aspect declaration has specific language in XML and tools are encapsulated into IDE plug-ins. The aspect oriented programming (AOP) instrumentation process modifies the Java byte code to add runtime hooks around point cuts into agent definition. Those hooks collect reflection information and invoke advices. But to be really dynamic, aspect weaver should be launched by program, this means when a specific event occurs or after a given request. Few aspect tools have this ability. The JBoss AOP [15] instruments takes point cuts definition from an XML file or the metadata file or the annotation tags already embedded in the byte code by the annotation compiler. Now there are three different modes to run our aspect oriented applications: precompiled, load time or hot swap. Hot swap weaving is the most dynamic use when we need to enable aspects in runtime and don't want that the flow control of our classes be changed before that. When using this mode, our classes are instrumented a minimum necessary before getting loaded, without affecting the flow control. If any join point is intercepted in runtime due to a dynamic aspect operation, the affected classes are weaved, so that the added interceptors and aspects can be invoked. As the previous mode, hot swap has some drawbacks that need to be considered such as performance perturbation. But this has a real impact into a reactive environment and for the current case study, this feature is not so essential. Also, our AOP choice has been to adopt JBoss AOP toolkit and associated IDE plug in. Figure 1: Interaction diagram between main parts of a mobile agent system 2.2 Mobile aspect declaration The first step in creating a mobile aspect in JBoss AOP is to encapsulate the whole mobile feature in a Java class. The kernel of the mechanism is based on use of distributed services. The first one is a lookup service which is used to register not only mobile agents but also mobile agent hosts. On our deployment diagram, there is one such registry per node of the network. An agent host is a candidate to a future reception of a mobile agent. It has to publish its reference into the registry of the node where it is. After that, it will be accessible by a mobile agent. Agent server is first a server which creates and manages mobile agents. After creation, mobile agents are published into local registry (1). When a mission is available, mobile agent can book it and then this agent can start to realize it. A mission consists in two parts: a route of nodes and a set of actions. The route is a sequence of computers which support agent host. The set of actions is written in an extern piece of code. Also to find out first node, mobile agent has to look up it into the lookup service. We can identify two aspect definitions: one for mobile agent, another for agent host. The role of the first mobile agent aspect covers the initial part of the mobile agent life: from its creation until its first publication. When mobile agent starts its mission, its configuration is done by the application of aspect called MobileAspect. The business logic of mobile agent is similar to an automaton with six states: first reading the mission, preparing migration, migration, negotiation, application of the mission, updating it description into registry. A second aspect is about authorized actions on an agent host. This one has also the role of a gate keeper. This means that it has to check what a mobile agent is going to do before operating its mission. Also, this aspect called HostAspect, is coupled with another one called SafetyAH, (figure 2). HostAspect is used to management of a set of mobile agents. An agent host knows what kinds of mobile agent it is waiting for, also this aspect implements this control before checking the rights of a permitted incoming agent. It business logic is simply an automaton with four states: waiting for an incoming message, negotiation, observation of activity, sending output message. Because an agent host has to be published into lookup service, it supports also first aspect MAAspect. When first part of its mission is ended, mobile agent updates own information in the local lookup service. This will be useful for administration of mobile agents. Next, it continues until the end of its mission. 2.3 Safety aspect declaration When a mobile agent is imported by an agent host, it cannot start its local activity before checking by host whether or not the permissions of mobile agent are

25 5 enough. Because this control algorithm is similar for all distributed system. Permission management is intrinsically mobile agent importation, we defined a new aspect called SafetyMA for the creation of negotiation step. During this step, mobile agent has to send its requirements to agent host; it means all local resources used for by agent with action. Its requirements can not be computed at the compile time, but at its entrance. This corresponds to agent loading time on agent host. A dynamic weaver can respect this strategy. Another aspect is applied to agent host which contains a definition of permissions with signature and localization. During its negotiation step (figure 2), agent host receives safety request from mobile agent. Its business logic starts by the analysis of this demand (figure 3) and the comparison with the accepted operation on the host and also permission list. This aspect called SafetyAH, creates from these collections, an AccessController instance which realized the controls/ If the request is satisfied, mobile agent could operate its local activity on the host. If the demand is not satisfied, host rejects the request and raises an exception. When negotiation step is satisfied, a security manager is created by agent host to control local activity. This last element is essential when mobile agent executes a local script. The actions which belong to that script have to respect initial contract of mobile agent with agent host. The lifetime of an AccessController instance depends on the safety request, but a SecurityManager instance observes mobile agent until the end of its mission. Then it will notify agent host and built a report about its own activity. It could be used for a post mortem analysis of mobile agent system. In previous section, two aspect were defined which can be used at compile time. But this second subset of aspects can not be used before deployment of all the Figure 2: State chart of two main piece of mobile agent system dynamic, for instance, when a mobile agent has realized an operation on a main agent host, it can operate on all hosts which depend on the main one. At the opposite, if a local resource is not accessible on a specific kind of agent host, then this property can be kept to simplify negotiation step on next host. But the scope of our safety aspects allows us to extract all safety property control from the source code of our project. Now, we can evolve separately the business code of distributed system (work of developer) and safety concerns which are managed by administrator or architect of a project. 2.4 Instrumentation aspect declaration Instrumentation is a large spectrum of activities from deployment step to runtime observation (performance, security, transaction state, etc). At the first level of domain is use of log files. We decided to manage centralized log information. But, if the used of log file is a basic example in all AOP framework, it becomes more complex into a distributed system. We want not only log centralization but also log consolidation. Standard UNIX syslogd offers UDP-based log forwarding to a central log consolidator today. We need additional features that make it a powerful tool for log forwarding, log centralization and log consolidation. We decided to use a technical framework, called JMX (for Java Management extension [16]. JMX specification defines instrumentation of services as MXBeans, agent architecture and standard services. The contract for MXBeans is simple, easy to implement, and unobtrusive for managed resources. Furthermore, the architecture set in the specification decouples the management clients from the managed resources,

26 6 increasing the reusability of JMX-based components. Also, its creation is candidate to an aspect definition. Centralized log consolidation offers the following benefits: easier log file analysis, increased security, simplified archiving of logs. A centralized log provides a single location for the administrator to perform log file analysis. It offers a single view of events that impact mobile agent systems. A security breach might compromise the local logs but not the centralized copy. Moreover, it is usually simpler to archive a set of centralized logs rather than per-system logs. Using the JMX technology, a given agent host is instrumented by one remote objects known as Managed Beans, or MBeans. These MBeans are registered in a coremanaged object server, known as an MBean server. The MBean server acts as a management agent. The JMX technology provides scalable, dynamic management architecture. Every JMX agent service is an independent module that can be plugged into the management agent, depending on the requirements We defined standard connectors (known as JMX connectors) that enable us to access JMX agents from remote management server. JMX connectors using different protocols provide the same management interface. Figure 3: Sequence diagram of negotiation step. Consequently, a management application can manage resources transparently. We defined a third subset of aspects which contains a definition for the creation of the instrumentation classes for agent host. This one uses it as a local logger but the logger exposes its interface on Jini protocol. Also, it is now possible to observe it from the agent server. Now, administrators are able to know where mobile agents are and also where are current problems, for instance negotiation failure, resource access violation. From agent host definition, aspect extracts location information and then creates and registers an instance of Instrumentation class. This instance is launched as a parallel flow of agent host. Its starter will be decided at run time. This means that the aspect weaver has to be call at run time. This last subset completes our application of aspect onto mobile agent system. 3. Aspect oriented design to aspect development Software engineering of mobile agent systems involves a number of concerns, including migration, safety, instrumentation, but also error handling, and a lot

27 7 3.1 Analysis modeling Figure 4: Class diagram for instrumentation of an agent host of other facets. The modeling, design, and implementation of many of these concerns are essential because they are inherently crosscutting as the system complexity increases. There is a pressing need for specifying aspect approach and its relations towards object-oriented design. This projection is also interesting because, we have already projection from mobile agent model towards object oriented specification. Some works are already published about mapping between AO design and OO design [17], this work is based on a UML specification of all the concepts belonging to aspect oriented programming and evaluation. Main disadvantage is lack generality, a specification is also linked to a aspect framework such as AspectJ or Spring AOP. A specification language like UML is a language specification providing a common interface usable for defining semantics applicable toward arbitrary AOP framework binding Suzuki and Yamamoto [18] propose a general approach to express aspects during design. They suggest an extension to UML to support aspects appropriately without violation to the meta model specification. For this purpose, they add to the meta model new elements for the aspect and the weaver and an existing element is reused to model the relationship class-aspect. We emphasize the generation of technology independent models using the Unified Modeling Language (UML) at different points in the software development lifecycle, e.g. requirements modeling, system analysis modeling, and design modeling with use of annotation. It helps to minimize changes on logic specification. We adopted UML notation with additional stereotype to place our aspects as sub classes of technical class belonging into JBoss AOP framework. Link between aspect and class is done by the use of a class association. For a couple aspects, class, this association represents the point cut for the weaving. Advice concept is described on Fig5 as a subclass of Interceptor class. This class belongs to JBoss, framework, this is why it is placed into its technical package. It is behavior that can be inserted between a caller and a callee, a method invoker and the actual method: for instance, between mobile agent and agent server. These aspect construction allow us to define cross-cutting behavior Point cuts tell the JBoss AOP framework which interceptors to bind to which classes, what metadata to apply to which classes, for example MobilePointcut is the link between MobileAgent class and one of its facets called MobileAspect. Aspect definitions of the three subsets are quite similar except for their business logic. The aspects MobileAspect and HostAspect are written with the use of Jini toolkit [19]. Jini is a simple set of Java Classes and services that allows node on a network (e.g., workstation) and agent (e.g., mobile collector) to access each other seamlessly, adapt to a continually changing environment, The package aspect.safety contains aspects and point cuts about security control. The aspect definition (SafetyAM and SafetyAH) are based on JCE [20] for the cryptographic features (used for certificate management) and JAAS [21] for authentication and authorization strategy. It allows us to plug an external authentication mechanism into message queue. The package aspect.instrumentation contains aspects definition for remote observation of mobile agent and agent hosts. These definitions are written by the use of JMX framework as remote protocol, and Java Platform Debugger Architecture (JPDA) [22]. It is a programming interface used by development and monitoring tools. It provides both a way to inspect the state and to control the execution of applications running in the Java virtual machine Design modeling In previous section, declarative description of our approach is explained. Now, we focus on precise definition of one mobile aspect. This aspect encapsulates technical facet of agent: migration via a specific protocol. It allows us to layer, rather than embed, functionality so that code is more readable and easier to maintain. When our migration mechanism will change, modification will be clearly identified.

28 8 Figure 5: Class diagram of first subset of aspect definition The point cut called MobilePointcut, defines an event linked to MobileAgent constructor call. So, at the construction, the central mechanism of a Jini system (the lookup service) is called and mobile agent is registered. Then this mobile agent is a mobile service available on the network. While registering, the mobile agent provides a callable interface to access its functionality and attributes those may be useful while querying the lookup from an agent host. Now, mobile agent is waiting until a mission is available. A new available mission is also an event which can be described by a point cut expression language. Also, it is a trigger for the preparation of migration. A mobile agent is just a temporary worker and its behavior is intrinsically asynchronous. We use another skill of JBoss server: it is a message queue server. Also, we defined a queue per group of mobile agents. The server notifies a MissionAdapter instance (Fig6), then this instance assigns mission to mobile agent. To add this filtering functionality, we modify MobileAspect and inject a pre statement; this is "Preparation migration" state (Fig2). The actions of mobile agent are basic: first, the access to agent host list. It involves discovering lookup, querying it for the specific agent host service (called acceptance) and invoking the callable interface of the service required. The callable interfaces are exposed and accessed There is symmetry with aspect HostAspect. When an agent host is plugged into the network, it locates the lookup service (by multi cast discovery) and registers its acceptance service there. Easily, Jini framework allows building up clusters of agents that know about one another and cooperate, creating a "federation" of agents. 4. Distributed audit Based on previous aspect approach of mobile agents, we defined a distributed monitoring application called DMA. Its main objective is the observation of agent host and mobile agent traffic. Main concept is a centralized log about distributed observations. 4.1 Mobility as a principle Data collection is a reference example for mobile agent system. In our context, we need not only to collect data but also to configure local collect algorithm. As mentioned previously, we use JMX framework for management of the distributed collection, but JPDA for local observation. The scheduling of our application has two main phases. First, there is a transient step where observer agents are deployed over network. Then, during a stable step, data are collected and consolidated on server;

29 9 Transient step consists in two waves. A first wave The option cpu is about local activities of mobile agents. This contains all resource access, but also the place where a security manager checks permissions during execution. Moreover, for each resource access, there is trace of mobile agent responsible and also local time stamp. Figure6: class diagram for mission reception exports a mobile agent, called Observer, on each agent host where data have to be collected. This kind of agent is just a mobile agent which executes an Observation mission. A second wave is configuration of all the observers. Several features can be set: data format, kind of observations (memory, cpu or network). Each value represents a set of events, for instance network is a filter for observing incoming agent and out coming agent. By the end of this wave, the stable step starts. This means that data collection starts by the use of mobile agent also (called Messenger). This kind of agent is just a mobile agent which executes an Messenger mission. This step is also structure as a loop where local data are first recorded on the agent host (into XML file) and then exported by a Messenger agent towards AgentServer. Messenger agent is able to encode and decode data via a specific format. AgentServer receives all Messenger agents, reads their data and consolidate all information into a single log file. Because the size of such file is limited, log file is sampled; each part is identified by a timestamp. 4.2 Monitoring information Because data format can be distinct, Messenger agent are essential to transform data into intermediate format of Agent Server. There are three kind of message depending on the previous configuration. If "network" is selected, migration information is sent. This is result of negotiation, duration of the presence of each mobile agent When memory is chosen, heap memory information is saved; the format follows the generation structure of the memory. Information about garbage collector is also saved (algorithm, frequency, etc). 5 Results Our approach was validated with a small group of ten nodes. It allows us to gather data for improving administration. Format of exchanged data is XML, but each agent host lays its own format, expressed an XML schema descriptor (XSD). Because, server format is unique, it transforms all input message into its server through an XSL transformation. 5.1 Data collection Fig7: transient step of DMA Each message is a single XML block of text with an associated namespace and a set of data. Main tag has a "facility" attribute. The facility can be thought of as a category that depends upon the mobile agent from which the message originates. A short example is given below; this message is exported by the end of configuration of an Observer agent: <?xml version="1.0" encoding="utf-8"?> <ns:message id="th1" facility="obs1" mnemonic="cpu" severity="2" time=" T12:10:21"> <ns:text value="end configuration"/> </ns:message> This message is a part of a larger discussion, called "th1" between observer and agent server. It marks the starter of data collection about local activities of agent host ("cpu" collection). During stable phase, a lot of packets can be received and it is necessary to filter input data and provide a time interval for collection. For instance, the given period is 20 ms and only trace of "allocate" events. These changes are done from the server to the observer by the use of JMX service. Then, the data format evolves to transmit only useful information.

30 10 <?xml version="1.0" encoding="utf-8"?> <ns:message id="th1" facility="obs1" mnemonic="cpu" severity="2" time=" T14:15:56"> <ns:cpu rank="1" self="81.17" accum="81.17" count="221010" trace="101152" method="dataaccessreader"/> </ns:message> In that case, the information is a snapshot of the current observation. The trace number seen above is related to stack traces in the file itself. This printout shows the rank, the amount of CPU consumed by that method call and total across the application execution, and then finally how many times that individual method was invoked. Then, it references back to a trace for that method invocation. Because this message is a member of a global thread of message, it is possible to rebuild its evolution over a period of time. 5.2 Benefits Fig8: data collection with mobile messengers. Upon initial setup of a new agent host, an observer is exported and messages are stored locally at the beginning. Then, automatically messages are routed via "Messenger" to a centralized location. The benefits are facilities to analyze what may have happened (normal behavior versus strange event) and simplification to archive collected logs off-line to removable media. Because cross format is XML, it is easy to filter a part of its contain to extract anomalies for instance, or to filter event that cost much more time than the others (the top 10 for instance). One of the main benefits of XML is that it separates data from its presentation. However, because we combine XML data with an XSL Transformations (XSLT) style sheet, we then have a powerful way to dynamically transform and present information in any format we want. Furthermore, often the structure of our XML stream created by our application does not match the structure required by other application parts to process that XML data. To transform the existing XML data structure into one that can be processed, we need to use XSLT. Moreover, transformation can provide SVG or GraphML representations that highlight cpu ratio or call numbers. Since the mechanics of applying XSLT style sheets to XML in Java code are generally the same, the process can be refactored out of the business-specific code into something more reusable. The chain starts with a source XML stream (though not necessarily a file), and applies a series of XSLT style sheets to it until it produces the final stream used to observe local activities. For instance, an XML stream describes memory management. It is the result of data collection (done by a mobile agent) about garbage collector activity on agent host. The Java runtime uses a garbage collector that reclaims the memory occupied by an object once it determines that object is no longer accessible. This automatic process makes it safe to throw away unneeded object references because the garbage collector does not collect the object if it is still needed elsewhere. Therefore, in agent host, the act of letting go of unneeded references never runs the risk of deallocating memory prematurely. This event is serialized into XML stream when mobile agent is physically on this agent host. At the end of its trip, a mobile agent contains a large data set and XSLT transformations are used to extract into specific order memory information. Transformations are about size of collected data or collection time or agent host address. It is also a strategy to compute memory amount used at a given time of a distributed computation. 4. Conclusions Aspect-oriented programming is a powerful new tool for software development especially mobile agent system. With JBoss AOP, we can implement your own interceptors, metadata, to make our mobile agent development process more dynamic and fluid. According to our experience, there was a number of crosscutting mobile agent system concerns which aspectoriented abstractions succeeded to cope with their modularization. This was often the case for mobility. For these agent properties, the design and implementation have shown expressive improvements in terms of separation of concerns. By the end, our objective of mobile agent instrumentation is achieved, especially for information collection for monitoring. References 1. Ivan Kiselev, Aspect-Oriented Programming with AspectJ, 2001, Sams, ISBN: Pavel Vrba, E.Cortese, F. Quarta, G. Vitaglione, Scalability and Performance of the JADE Message Transport System. Analysis of suitability for Holonic Manufacturing Systems.

31 11 this number of EXP. LNCS, vol. 4128, pp Springer, Heidelberg (2006) 3. Mobile agents, a.k.a distributed agents, according to Tryllian, doi: /ic: , IEE Seminar Mobile Agents - Where Are They Going (2001/150) London, UK, 11 April 2001, 4. A. Corradi, R. Montanari, C. Stefanelli, Mobile Agents Protection in the Internet Environment, Proceedings of the COMPSAC'99, IEEE Computer Society Press, Phoenix, October ' Jean-Paul Arcangeli, Vincent Hennebert, Sébastien Leriche, Frédéric Migeon, Marc Pantel. JavAct : principes, installation, utilisation et développement d'applications. Rapport de recherche, IRIT/ R, IRIT, février E. Spafford and D. Zamboni. A framework and prototype for a distributed intrusion detection system. Technical Report 98-06, COAST Laboratory, Purdue University, West Lafayette, IN , May Cyril Dumont, Fabrice Mourlin: Space Based Architecture for Numerical Solving. CIMCA/IAWTIC/ISE 2008: Cyril Dumont, Fabrice Mourlin: A Mobile Computing Architecture for Numerical Simulation CoRR abs/ : (2007) 9. Cyril Dumont, Fabrice Mourlin: Adaptive runtime for numerical code, 8th ENIM IFAC International Conference of Modeling and Simulation Evaluation and optimization of innovative production systems of goods and services (2010) pp Pavel Avgustinov, Aske Simon Christensen, Laurie Hendren, Sascha Kuzins, Jennifer Lhot ak, Ond rej Lhot ak, Oege de Moor, Damien Sereni, Ganesh Sittampalam, and Julian Tibble. abc: An extensible AspectJ compiler. In AOSD, mar 2005, pages ACM Press. 11.Chitchyan, R. et al. Survey of Aspect-Oriented Analysis and Design. AOSD-Europe Project Deliverable No: AOSD- Europe-ULANC Garcia, A., Chavez, C., Kulesza, U., Lucena, C. The Role Aspect Pattern. Proc. of the 10th European Conf. on Pattern Languages of Programs (EuroPLoP 05), July 2005, Irsee, Germany. 13.Griswold, W. et al, "Modular Software Design with Crosscutting Interfaces", IEEE Software, Filman, R. et al. Aspect-Oriented Software Development. Addison-Wesley, Tom Marrs, Scott Davis, JBoss at Work: A Practical Guide, O'Reilly (2004) 16.Benjamin G. Sullins and Mark B. Whipple, JMX in Action, pages, ISBN: Kendall, E. "Role Model Designs and Implementations with Aspect-oriented Programming". OOPSLA 1999, pp Extending UML with Aspects: Aspect Support in the Design Phase. 3er Aspect-Oriented Programming (AOP) Workshop at ECOOP 99. Junichi Suzuki, Yoshikazu Yamamoto. 19. Scott Oaks, Henry Wong, Jini in a Nutshell, O'Reilly Media, March David Hook, Beginning Cryptography with Java, ISBN13: , ed. Wrox, Michael Cote, Java Authentication and Authorization Service (JAAS) in Action, ed Wiley, Harbourne-Thomas A., Bell J., Brown S., "online: Professional Java Servlets 2. 3", ISBN: X, 2003 Youssef Hannad is PhD student at Paris 12 University. His position is set by Team Up corporate. His subject is on aspect for mobile agent management. Fabrice Mourlin. Is associated professor at Paris 12 University since He manages a working group on mobile agent and space computing. He obtained HDR habilitation in 2008 at Paris 12 university. Current projects are network monitoring and numerical computing.

32 12 A Modified Algorithm of Bare Bones Particle Swarm Optimization Horng-I Hsieh 1 and Tian-Shyug Lee 2* 1 Graduate Institute of Business Administration, Fu-Jen Catholic University Hsin-Chuang, Taipei County 24205, Taiwan, ROC 2 Department of Business Administration, Fu-Jen Catholic University Hsin-Chuang, Taipei County 24205, Taiwan, ROC Abstract Bare bones particle swarm optimization (PSO) greatly simplifies the particles swarm by stripping away the velocity rule, but performance seems not good as canonical one in some test problems. Some studies try to replace the sampling distribution to improve the performance, but there are some problems in the algorithm itself. This paper proposes a modified algorithm to solve these problems. In addition to some benchmark test functions, we also conducted an application of real-world time series forecasting with support vector regression to evaluate the performance of the proposed PSO algorithm. The results indicate that the modified bare bones particle swarm optimization can be an efficient alternative due to the smaller confidence intervals and fast convergence characteristics. Keywords: Heuristic Optimization, Particle Swarm Optimization, Bare Bones PSO, Support Vector Regression, Time Series Forecasting. *Corresponding author. 1. Introduction Particle swarm optimization (PSO) is a population-based heuristic method developed by Kennedy and Eberhart in 1995 [1]. The PSO algorithm is inspired by the collective motion of biological organisms, such as bird flocking and fish schooling, to simulate the seeking behavior to a food source. A PSO algorithm is initialized with a population of random particles treated as a point in D-dimensional search space. To find the optimum solution, each particle adjusts the direction through the best experience which it has found (pbest) and the best experience been found by all other members (gbest). Therefore, the particles fly around in a multidimensional space towards the better area over the search process. The PSO system initially has a population of random solutions and then searches for optimum solution by updating process. Each particle consists of three vectors: the position for ith individual particle x i = (x i1, x i2,..., x id ), the best previous position that the ith particle has found p i = (p i1, p i2,..., p id ), and its velocity v i = (v i1, v i2,..., v id ). The performance of each particle is measured using a fitness function varying from problem in hand. During the iterative procedure, the particle s velocity and position are updated by old old vid c1 rnd() ( pid xid ) new vid (1) old c2 Rnd() ( pgd xid ) new old new x x v (2) id id id where c 1 and c 2 are two positive acceleration constants, χ is a constriction factor, p id is the pbest of ith particle and p gd is the gbest of the group, and rnd() and Rnd() are two random numbers uniformly generated from [0,1]. In a PSO system, particles change their positions at each time step until a relatively unchanging position has been encountered or a maximum number of iterations has been met. Kennedy [2] proposed a new PSO where the usual velocity formula is removed and replaced with samples from a Gaussian distribution. The velocity-free bare bones (BB) PSO was inspired by the observation that histogram sampled by the canonical particle swarm is appeared to be normally distributed around (p id + p gd ) / 2, with a standard deviation of p id p gd. Many factors have been found to determine how successful the problem-solving process are often problem dependent. The bare bones PSO using information drew from a Gaussian distribution greatly simplifies the particles swarm algorithm, but the performance in Gaussian version is not as good as canonical PSO [3][4]. Some researchers [3][5] try to examine the bell shape distribution and replace the Gaussian random number generator by the appropriate one to reproduce the behavior of canonical algorithm, and improvements have been observed. However, the problem

33 13 of bare bones PSO might not be the replaced distribution, but the algorithm itself. The bare bones PSO might suffer from premature convergence or converge to a point neither global nor local optimum. The aim of this paper is to propose a new algorithm to avoid the problem mentioned above. Furthermore, the performance of the proposed algorithm in some benchmark functions and a real-world application are investigated. The remainder of the paper is organized as follows: An overview and a closer examination on convergence behavior of bare bones PSO is given in Section 2. Section 3 provides the modified algorithm of bare bones particle swarm optimization. Benchmark functions to measure the performance of the different approaches are provided in Section 4. Section 5 presents the results of a real-world application and conclusions are drawn in Section Bare Bones PSO There are several problems appear in the bare bones PSO. Firstly, the best particle in bare bones acts different from the canonical one. In a canonical particle swarm, particles change their positions at each time step. On the other hand, the best particle of the neighborhood in bare bones PSO simply stands in its best previous position due to the random number generator definition. If the other particles move too close to the best one in their neighborhood, the particles may converge on a point that is neither the global nor the local optimum. Consider a simplified situation shown in Figure 1, which with only two particles and one dimension. Since particle A is the best one, it will always stand in the same position as particle B moves to the right side without finding out any better result. If particle B flies into [P A, P B ] or [0, P A ] but too close to particle A, the PSO system will become inactive. The particles will move very slowly in future iterations; even converge on a wrong position. Fortunately, as the number of particles grows, the probability that system becomes ineffective decreases. However, if the problem becomes complicated, determining an appropriate number of particles might be a difficult task because the growing size of swarm also increases the time of computation. Y Additionally, bare bones particle swarm might suffer from premature convergence. When the best particle locates in suboptimal position, it tends to mislead all the other particles to get stuck in this local optimum. Each particle in the neighborhood can fast approach to the best area within few iterations by making large step sizes even if the particles whose personal previous best position are far away from the global best position. Consider the simple situation depicted in Figure 2. All the particles move fast toward to the p g inside a local optimum area. If each of the particles fails to hit the region of the global optimum, the whole system might lose exploration capability as the standard deviation cannot be back to large value. D j Y 0 0 Global Optimum P A Fig. 2 All of particles are prematurely converging to local optimum. 3. A Modified Bare Bones PSO Local Optimum As mentioned in the preceding section, the best particle without momentum might harm the performance of the whole system. Our strategy is to modify formula of standard deviation by adding a new parameter. If the particle is the best one in its neighborhood, the new standard deviation is computed as P A P g P g D i D i δ p id p gd = δ p gd p gd, (3) 0 D P A P B Fig. 1 A simplified particle swarm system. where δ is a constant, which may be a number either larger or small than 1. The offset of the best position behaves similar to the momentum term in velocity update rule of canonical PSO. Table 1 shows the results of the proposed

34 14 standard deviation computing strategy with δ = 1.2. As the results reveal, the new strategy can efficiently solve the above-mentioned problem. Table 1: Mean results of De Jong after 50 trials of 3000 iterations Dimensions / Number of particles Methods original modified 2.6E E original modified 4.93E E-103 5E E-189 Numbers that are less than E-324 are rounded to 0. Since the best particle might mislead all the other particles into local optimum very fast, one thought is to slow down the speed of the movement. Because of forcing the particles to reach the region close to the global optimum might reduce the chance of getting stuck in local optimum; a slower convergence by using a smaller step size in the earlier stage might be beneficial. On the other hand, if the particle inside one of a local optimum leading by the best particle also has the capability to reach the region near the global optimum, it is likely to escape from the inferior suboptimal. Thus, there are two strategies can deal with the premature convergence situation: 1. Offset the mean to force the particle not being very close to the best particle during the early stage. 2. Slow down the biased exploration by constraining the standard deviation and offsetting the mean together during the early stage. Thus, the mean and standard deviation of the Gaussian distribution used to update the position in the early stage becomes P P 2 id gd XiD ~ N1, 2 PiD PgD where ω 1 and ω 2 are two constriction parameters which can be either a fix or dynamic value over the early stage. A series of experiments were conducted by using eight benchmark functions to investigate these ideas. All the test functions were implemented in 30 dimensions except for the two-dimensional Schaffer s f6 function. Their definitions and initial range are shown in Table 2. Twenty particles were used in the test presented here. Each experiment was implemented 100 times for 3000 iterations. All algorithms are initialized asymmetrically with the ranges as shown in Table 2. (4) Table 2: Benchmark functions Name Equation D Initialization De Jong Schwefel 1.2 Schaffer s D 2 x i 1 i D i i1 j1 f ( x ) 30 (50,100) D 2 f ( x ) ( x j ) 30 (50,100) D 2 2 (sin x y ) f ( x ) (50,100) D f x y Rosenbrock D 2 2 f( ) 100x 1 1 i 1 i xi x i Rastrigrin f D 2 ( ) cos2 10 i 1 i i Schwefel 1 2 x 30 (15,30) D x 30 (2.56,5.12) D D f ( ) x sin 2.6 i 1 i xi Griewank Ackley x 30 (-500,-250) D D D 1 2 xi f( x ) xi cos 1 30 (300,600) D 4000 i1 i1 i 1 D 2 f( x) 20exp 0.2 x i 1 i D 1 D exp cos2 x 20 i 1 i e D 30 (16,32) D Table 3 summarizes the results. In this test, ω 1 and ω 2 were set as constants with a large and a small value, and the constriction was only conducted before the 100th iteration. After the adjustment in the early stage, the ω 1 and ω 2 were set back to 1 to ensure the convergence. Also, the position offset factor of the best particle δ was set to 1.2. Each setting has good performance in some test problems are shown in bold. Due to the generalization ability consideration, the third setting which ω 1 and ω 2 are set to 0.7 will be used thereafter. For convenience, the rest of this paper will use the term BBM (modified bare bones) as the abbreviation. Table 3 Mean results of parameter test after 100 trials of 3000 iterations Schwefel Schaffer s De Jong Rosenbrock 1.2 f6 M 0.7 M 0.4 MS 0.7 MS E-58 (3.15E-57) 3.27E-112 (2.33E-111) 7.23E-70 (2.89E-69) 1.47E-118 (1.10E-117) ( ) ( ) ( ) 1.21E-06 (4.35E-06) 2.72E-03 (4.38E-03) 1.75E-03 (3.75E-03) (4.56E-03) 1.94E-04 (1.37E-03) ( ) ( ) ( ) ( ) Rastrigrin Schwefel 2.6 Griewank Ackley M ( ) (585.55) ( ) 1.70E-14 (4.75E-15) M (0) ( ) 0 (0) 9.95E-16 (6.09E-16) MS ( ) (428.13) 0 (0) 1.48E-14 (5.35E-15) MS (0) ( ) 0 (0) 1.53E-15 (1.37E-15) Note that M means only mean offset method is used, and MS means constrict the mean and standard deviation in the same time. Standard deviations are shown in parentheses. Numbers that are less than E-324 are rounded to 0.

35 15 4. Experimental Results This section compares the performance of BBM with BB and canonical PSO. The three coefficients of canonical PSO were set as χ = and c 1 = c 2 = 2.05 [6]. Eight benchmark functions shown in Table 2 were used to compare the performance of BBM with those of other algorithms. Gbest, Ring, and Square topologies were tested for all algorithms. A swarm size of 20 was used in all experiments, and each experiment was run 100 times for 3000 iterations. Also, the algorithms were initialized asymmetrically and the ranges did not contain global optimum, which can be found in Table 2. As seen in Table 4, BB shows comparable results with canonical PSO on some test functions, but has the worst result on Rosenbrock with all topologies. On the other hand, BBM shows the best results on 6 out of 8 functions across all topologies, and outperforms BB on 7 out of 8 functions when using the Square topology. Note that BBM is the only one able to find the global minimum to the Rastrigrin, Griewank and Ackley functions, and the best results with the smallest standard deviations to Rosenbrock function. Figure 3 shows the mean performance best over time with the Gbest topology. As seen in Figure 3, BBM achieved a faster reduction than BB on all of the test functions. In summary, BBM provides better results with smaller confidence intervals compared to BB, thus can be a competitive optimizer on these test functions. 5. Model Selections in Support Vector Regression Support vector machine (SVM) is a novel neural network algorithm based on statistical learning theory [7]. With introduction of Vapnik s ε-insensitivity loss function, the regression model of SVMs, called support vector regression (SVR), has been receiving increasing attention to solve nonlinear regression problems. In the modeling of SVR, one of the key problems is how to select model parameters correctly, which plays an important role in good generalization performance. However, no general guidelines are available to choose the free parameters of an SVR model. This section demonstrates a financial time series forecasting problem by using PSO to search the optimal parameters of SVR model selections. In order to evaluate the performance of the proposed approach, the Nikkei 225 closing cash index is used as the illustrative example. Table 4: Mean results of eight test functions after 100 trials of 3000 iterations De Jong Schwefel Schaffer 1.2 s f6 Rosenbrock Canonical Gbest 2.09E-23 (1.58E-22) 3.71E-16 (3.71E-15) 3.98E-03 (4.80E-03) ( ) Ring Square BB Gbest Ring Square BBM Gbest Ring Square Canonical Gbest Ring Square BB Gbest Ring Square BBM Gbest Ring Square 1.53E-23 (2.62E-23) 7.06E-30 (4.10E-29) 6.53E-23 (6.53E-22) 1.52E-10 (1.08E-09) 8.88E-20 (5.72E-19) 7.23E-70 (2.89E-69) 9.62E-48 (1.97E-47) 3.85E-53 (1.36E-52) Rastrigrin ( ) ( ) ( ) ( ) ( ) ( ) ( ) 0 (0) 0 (0) 5.62E-30 (3.19E-29) 1.19E-28 (3.34E-28) ( ) ( ) ( ) ( ) 3.72E-05 ( ) ( ) Schwefel (259.66) (195.36) (222.24) (336.97) (253.16) (268.43) (428.13) (432.08) (356.85) 2.60E-03 (4.30E-03) 1.55E-03 (3.58E-03) 5.44E-03 (4.85E-03) 1.30E-03 (3.14E-03) 1.63E-03 (3.63E-03) (4.56E-03) 2.14E-03 (4.05E-03) 1.17E-03 (3.17E-03) Griewank ( ) ( ) ( ) ( ) ( ) ( ) 0 (0) 0 (0) 0 (0) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) Ackley ( ) ( ) ( ) (4.29) (1.90) (5.28) 1.48E-14 (5.35E-15) 1.01E-14 (3.03E-15) 9.81E-15 (3.09E-15) Because the real optimum is unknown, a grid-search method is applied to find the best combination of parameters. The searching space of the three parameters was set in [2-15,2 15 ] with a step size 0.5. The minimum root mean squared error (RMSE) found by the girdsearch method is considered as optimum in forecasting the Nikkei closing cash index. The task of PSO here is to find a set of parameters with acceptable even better accuracy. The number of dimension is equal to 3 as there are three free parameters in SVR. Each trial was randomly initialized in [2-15,2 15 ]. Because modeling SVR is a timeconsuming task, 10 particles were used in this section, and each experiment was implemented 50 times for 200 iterations. Any trial reaches the criterion equal to was treated as a success case. Two constriction parameters in BBM were set to 0.7, and set back to 1 after 30 iterations.

36 De Jong 10 4 Schwefel 1.2 Canonical Canonical 10 0 BB BBM 10 2 BB BBM Schaffer's f6 Canonical BB BBM 10 9 Rosenbrock 10 8 Canonical BB BBM Rastrigrin Canonical BB BBM 10 3 Schwefel 2.6 Canonical BB BBM Griewank Canonical 10 2 Ackley Canonical 10 0 BB BBM 10 0 BB BBM Fig. 3 Comparison between three algorithms with eight benchmark functions.

37 17 Table 5 shows the results obtained from three algorithms. All algorithms using the Ring topology return better results than the Gbest one. Canonical PSO with Ring topology is the best one with the smallest mean RMSE. In addition, BBM using Ring topology outperforms BB in lower RMSE and the smallest standard deviations. Figure 4 shows that BBM converges very fast with both Gbest and Ring topologies. Furthermore, the strategy that a coarse grid followed by a finer grid using in SVM with two parameters might not be applicable here, because there are three parameters in SVR, the better solution is not necessarily nearby another good solution. Thus, the grid-search might not easily find any better solution. On the other hand, improvement is still possible because the early stop criterion with 200 iterations was used here. The minimum found by three algorithms shows in Table 5 also prove its potential. Table 5: SVR results of three algorithms Canonical BB BBM Gbest Ring Gbest Ring Gbest Ring Mean (0.54) (0.34) (0.85) (0.35) (0.53) (0.23) Max Min Successful rate 86.00% 96.00% 80.00% 92.00% 56.00% 94.00% 10 7 SVR Gbest Canonical SVR Ring BB BBM Canonical BB BBM 6. Conclusions This paper proposed a modified bare bones PSO by adding three extra parameters to correct the problems appears in original bare bones PSO. The modified algorithm shows advantages in better performance with smaller standard deviation and faster convergence characteristics. The restart strategy to deal with the problem converging to local optimum might harm than benefit the swarm. In Contrast, the offset strategy applied in early stage shows very good performance without harming the capability of convergence. Therefore the proposed algorithm can be a competitive optimizer. Acknowledgments This research was partially supported by the National Science Council of the Republic of China under Grant Number NSC E MY2. References [1] J. Kennedy, and R.C. Eberhart, Particle Swarm Optimization, in Proc. IEEE Intell. Conf. Neural Networks, 1995, Vol. IV, pp [2] J. Kennedy, Bare Bones Particle Swarms, in Proc. IEEE swarm Intell., 2003, pp ,. [3] J. Kennedy, Probability and Dynamics in the Particle Swarm, in Proc. IEEE Cong. Evolutionary Computation, 2004, Vol. 1, pp [4] J. Kennedy, Dynamic-probabilistic Particle Swarms, in Genetic Evolutionary Computation Conf., 2005, pp [5] T.J. Richer, and T.M. Blackwell, The Lévy Particle Swarm, in Proc. IEEE Cong. Evolutionary Computation, 2006, pp [6] M. Clerc, and J. Kennedy, The Particle Swarm: Explosion, Stability, and Convergence in a Multi-dimensional Complex Space, IEEE Trans. Evolutionary Computation, 2002, Vol. 6, pp [7] V.N. Vapnik, The Nature of Statistical Learning Theory, 2nd ed., NY: Springer, Fig. 4 SVR results of three algorithms.

38 18 Web Personalization of Indian e-commerce Websites using Classification Methodologies Agarwal Devendera 1, Tripathi S.P 2 and Singh J.B. 3 1 School of Computer Science & Information Technology, Shobhit University, Research Scholar Meerut, U.P , India 2 Department of Computer Science & Engineering, Gautam Buddh Technical University, IET Lucknow, U.P , India 3 School of Computer Science & Information Technology, Shobhit University Meerut, U.P , India Abstract The paper highlights the classification methodologies using Bayesian Rule for Indian e-commerce websites. It deals with generating cluster of users having fraudulent intentions. Secondly, it also focuses on Bayesian Ontology Requirement for efficient Possibilistic Outcomes. Keywords: E-commerce, Possibilistic Outcome, Bayesian Rule, Ontology. 1. Introduction Electronic Commerce is fast emerging as most popular method of purchasing, let it be a small pen drive or bulky LED TV. Recent survey [3] has estimated that around 3-5% of Indians have transacted or are well versed with working of online shopping websites. The strategy which is being followed until now related to the various policy initiatives like: Consumer Proportion: This model is being propagated by the government based on certain guidelines for the protection of consumers. Legality: It deals with formal recognition of electronic signatures; In India digital signatures are necessary for e-tendering. Security: Central Government has issued its policy relating to cryptography techniques to ensure secure electronic commerce in third party transfer. In order to deal with security and web personalization [2] issues we develop two basic classification methods: Naïve Bayes and K-nearest neighbor. 2. Our Model In order to make our model more illustrative we are taking example of Predicting Fraudulent Transaction. An Indian e-commerce company has a very large customer base; each customer has to submit his personal information before making a transaction. In this way each company is acting as a record and the response of internet is given as Z = {Fraudulent, Trustworthy} (1) these are the classification in which we can categorize a customer. By analyzing from a sample e-commerce site we are able to find out that in case of Fraudulent the customer-id should be reposted to the e-fraud cell. Two set of data are taken to check the consistency of data. Table 1: Report of Customer on e-commerce site. Reporting to e-fraud cell No Reporting Required Total fraudulent trustworth y Total Naïve Bayes In order to classify record into m classes by ignoring all predictor information X 1, X 2,.., X p is to classify the record as a member of majority class. For example in our case naïve rule would classify all the customers to be

39 19 Trustworthy, because 90% of the companies were found to be Truthful. Naïve Bayes classifier [1] is an advanced version of Naïve rule. The logic to introduce Bayes is to integrate the information given in the set of predictors into the naïve rule to obtain more accurate classifications. The methodology suggests in finding out the probability of record belonging to a certain class is evaluated on the prevalence of that class along with additional information that is being given on that record in terms of X information. Since our dataset is very large we prefer Naïve Bayes method. In a classification task our goal is to estimate the probability of membership to each class given a certain set of predictor variables. This type of probability is called a conditional probability. In our example we are interested in P (Fraudulent Reporting to e-fraud cell). In general, for a response of m classes C 1, C 2,.., C m and the predictors X 1, X 2,.., X p we compute as: P (C i X 1,,X p ) where i = 1, 2,, m. (2) When the predictors are all categorical we can use a pivot to estimate the confidential probabilities of class membership. Consider its application in our example we compute the probabilities divided into two classes as: For P (Fraudulent Reporting to e-fraudulent cell) = 20/120 and P (Trustworthy Reporting to e-fraudulent charges) = 100/120. The above statement indicates that although the firm is still more likely to be Trustworthy than Not Trustworthy, the probability of its being Truthful is much lower than the naïve rule. However, the method usually gives good result partly because what is important is not the exact probability estimate but the ranking for that case in comparison to others. In order to convert the desired probabilities into class probability we use Bayes Theorem. The Bayes Theorem gives us the following formula to compute the probability that the record belongs to class C i : P( X1,..., XP C1 ) P( C1 ) (3) P( Ci X1,..., XP) P( X1,..., XP C1 ) P( C1 )... P( X1,..., XP Cm) P( Cm ) C i : To compute the numerator we filter two pieces of information i) The proportion of each class in the population ii) [P(C 1 ) P(C m )] The probability of occurrence of the predictor vales X 1, X 2,, X p within each class from the training set. We develop another table of the User which is categorized as Frequent Buyers and Occasional Buyers, for each of these two categories of Buyers we have information on whether or not reporting has been done, and whether it turned out to be Fraudulent or Trustworthy. Reporting to e-fraud cell Table 2: Sample of 10 users. User-Type Status Yes Occasional Buyer Fraudulent No Occasional Buyer Trustworthy No Frequent Buyer Fraudulent No Frequent Buyer Trustworthy No Occasional Buyer Trustworthy No Occasional Buyer Trustworthy No Frequent Buyer Trustworthy Yes Occasional Buyer Fraudulent Yes Frequent Buyer Fraudulent No Frequent Buyer Fraudulent The probability of fraud can be defined by four possible states {Yes, Occasional Buyer}, {Yes, Frequent Buyer}, {No, Occasional Buyer}, {No, Frequent Buyer}. i) P(Fraudulent Reporting = Yes, Customer Type = Occasional Buyer) = 1/2 = 0.5 ii) P(Fraudulent Reporting = Yes, Customer Type = Frequent Buyer) = 2/2 = 1 iii) P(Fraudulent Reporting = No, Customer Type = Occasional Buyer) = 0/3 = 0 iv) P(Fraudulent Reporting = No, Customer Type = Frequent Buyer) = 1/3 = 0.33 We can extend this for Naïve Bayes probabilities, for analyzing the conditional probabilities of fraudulent behavior Reporting to e-fraud cell = Yes, and User Type = Occasional Buyer, the numerator is a proportion of Reporting to e-fraud cell. Instances amongst the type of Buyers, times the proportion of Fraudulent Customers = (3/4) (1/4) (4/10) = To get the actual probability we calculate the numerator for the conditional probability of truth given Reporting to e-fraudulent Cell = Yes; Type of Customer = Occasional Buyer; The denominator is then the sum of two conditional probabilities = ( ) = 0.14 Therefore the conditional probability of fraudulent behaviors is given by P NB (Fraudulent Reporting to e-fraudulent cell = Yes; Buyer Type = Occasional ) = (3/4)(1/4)(4/10) (3/4)(1/4)(4/10)+(1/6)(4/6)(6/10)

40 20 = 0.075/0.14 = 0.53 P NB (Fraudulent Reporting to e-fraudulent cell = Yes; Buyer Type = Frequent ) = P NB (Fraudulent Reporting to e-fraudulent cell = Yes; Buyer Type = Occasional ) = Rank Ordering of probabilities are even closer to exact Bayes method than are the probabilities themselves, to further analyze we can use classification matrix. 2.1 Advantages & Disadvantages of Naïve Bayes Classifier The logic of using Naïve Bayes Classification Technique [7] is to attain computational efficiency and good performance. 2.2 Fuzzy Information Classification and Retrieval Model The above section deals with a classification technique [6] by which we can categorize the customer visiting our site based on their transaction history. In this section we have highlighted the problem which our customer face while selecting the best possible combinations of product, the problem is because of the uncertainty in Semantic Web Taxonomies [8]. Consider Indiatimes shopping portal shown in fig. 1. Fig. 1: Indiatimes Shopping Portal. If a buyer wants a laptop in the range of Rs < x < Rs.35000, and with features F = {f1, f2, f3} in brands B = {b1, b2}, then he must be shown the best possibilistic outcome of the above query. The above problem looks very simple but it is not so, there exists an uncertainty in the query, what if, if there is no laptop with all the features of F present in Brand B. Here comes a probabilistic method to overcome such situation. In our method, degrees of subsumption will be covered by Bayesian Network based Ontology s [4]. The Venn diagram shown in figure 2 f1 f2 f3 Price Range - I Price Range - II Laptops Electronic Items Fig. 2: Venn Diagram Illustrating Electronic Items with Laptops as one of their Categories & their Overlap. Our method enables the representation of overlap between a selected concept and every other is referred taxonomy. The Price Range-I represent the prices at the start of the price band while Price Range-II represent the higher side of the price band. The overlap is logic term expressed as Selected Referred Overlap [0,1] (4) Referred The overlap region represents the value 0 for disjoint concepts and 1, if the referred concept is subsumed by the selected one. This overlap value can be used in information retrieval tasks. The match with the query is generalized by the probabilistic sense and the hit list can be sorted into the order of relevance accordingly. If F and B are sets; then F must be in one of the following relationships to B. i) F is a subset of B ie F B. ii) F partially overlaps B ie x, y : ( x F x B) ( y F y B) iii) F is disjoint from B ie F B Based on these relations we develop a simple transformation algorithm. The algorithm processes the overlap graph G in a Breadth First manner starting from root concept defined as CON. Each processed concept CON is written as the part of Solid Path Structure (SPS). The overlap values O for a elected concept F and a referred concept B B1 B2 B3 Mutual Overlap at B2 Brand

41 21 if else end F subsumes B then O := 1 C = F S B S if C = φ then O : = 0 else Σ m(c) c C O: = m(b) end Institutions for their financial support in his maiden endeavor. Raw Ontology File Structure Best Possible Outcome Refinement Stage Naïve Bayesian Transformation Quantifier Fig. 3: Computing the Overlap. If F is the selected concept and B is referred one, then the overlap value 0 can be interpreted as the conditional probability P(B' true F' true) S(F) s(b) 0 s (B) where S(F) and S(B) are taken is and interpreted as a probability space, and the elements of the sets are not interpreted as elementary outcomes of some random phenomenon. The implementation stages of the probabilistic search starts with the Input of Ontology Rule which are refined in Refinement Stage. It is than passed to the Quantifier which develops a set of Association Rules. It is then fed to the further preprocessing by the Naïve Bayesian Transformation module which finally generates the best possible overlapping outcome as shown in figure Conclusions & Future Scope The model in both the cases uses interactive query refinement mechanism to help to find the most appropriate query terms. The Ontology is organized according to narrower term relations. We have developed an algorithm in which taxonomies can be constructed without virtually any knowledge of Probability and Bayesian network. The future extension could be to expand it using Fuzzy Regression [7] with Bayesian Network. Acknowledgments Agarwal, Devendera thanks Darbari, Manuj without his essential guidance this research paper would not have been possible and also to management of Goel Group of Fig. 4: Implementation Framework. References [1] Bergenti, F., "Improving UML Designs using Automatic Design Pattern Detection", in 12 th International Conference on Software Engineering and Knowledge Engineering (SEKE), [2] Chen, Ming-Chung., Mining User Progressive User Behavior for E-Commerce using Virtual Reality Technique, M.S. Thesis, Faculty of Graduate School, University of Missouri-Columbia, [3] IAMAI, I-CUBE , Report by IMRB International, India, [4] Ding, Z., "A Probabilistic Extension to Ontology Language OWL", in 12 th Hawaii International Conference on Systems Science, [5] GuT, A Bayesian Approach for Dealing with Uncertain Concepts, in Conference Advances in Pervasive Computing, Austria, [6] Schafer, J.Ben., et. al., E-Commerce Recommendation Applications, as Grouplens Research Project, University of Minnesota, [7] Wang, Lipo., Fuzzy Systems & Knowledge Discovery, Springer, [8] Zongmin, Ma., Soft Computing in Ontologies and and Semantic Web, Springer, Agarwal, Devendera is currently working as Prof. & Director (Academics) at Goel Institute of Technology & Management, Lucknow. He has over 12 years of teaching & 5 years of industrial experience. Having done his B.Tech in Computer Science from Mangalore University in 1993, M.Tech from U.P.Technical University, Lucknow in 2006, he is pursuing his Ph.D. from Shobhit University, Meerut. Tripathi, S.P. (Dr.) is currently working as Assistant Professor in Department of Computer Science & Engineering at I.E.T. Lucknow. He has over 28 years of experience. He has published numbers of papers in referred National Journals. Singh, J.B. (Dr.) is currently working as Dean Students Welfare at Shobhit University, Meerut. He has 38 years of teaching experience and has published number of papers in referred National Journals.

42 22 High Accuracy Myanmar Handwritten Character Recognition using Hybrid approach through MICR and Neural Network Dr. Yadana Thein, San Su Su Yee University of Computer Studies Yangon (UCSY) Yangon, Myanmar Abstract This paper contributes an effective recognition approach for Myanmar Handwritten Characters. In this article, Hybrid approach use ICR and OCR recognition through MICR (Myanmar Intelligent Character Recognition) and backpropagation neural network. MICR is one kind of ICR. It composed of statistical/semantic information and final decision is made by voting system. In Hybrid approach, the features of statistical and semantic information of MICR have been used in back-propagation neural network as input nodes. So it needs a few input nodes to use. The back-propagation algorithm has been used to train the feed-forward neural network and adjustment of weights to require the desired output. The purpose of Hybrid approach to achieve the high accuracy rates and very fast recognition rate compare with other recognition systems. The experiments were carried out on 1000 words samples of different writer. Using Hybrid approach, over-all recognition accuracy of 95% was obtained. Keywords: MICR (Myanmar Intelligent Character Recognition), Hybrid approach, Feature of Statistical and Semantic Information, Back-propagation Neural Network, ICR, OCR. 1. Introduction Handwritten character recognition is the process of classifying written characters into appropriate types based on the features extracted from each character. It can be performed either online or offline. The researches on the recognition of the handwritten writing tend is a difficult task because of the differences of handwritings and of the irregularity of the writing of the same writer. Maybe among difficult tasks of handwritten character recognition, it is easier to recognize English character than Myanmar character. There are two main types of character recognition methods: Optical Character Recognition (OCR) and Intelligent Character Recognition (ICR). OCR is prospered in machine printed (i.e. typed) character recognition field. It translated the digitized images of text into machine-readable format. In handwritten character recognition field, OCR occur errors such as misrecognition, inconvenient, etc. ICR can successfully overcome these problems. ICR is an advanced version of OCR, which is used to enhance the accuracy in recognition levels. Myanmar language is the official language and widely used in many Myanmar states. In many offices were used the Myanmar language such as passport, bank, sales tax, railway, embassy, etc. So, it is a very importance to develop the high accuracy character recognition system for Myanmar language. Texts in the Myanmar language use the Myanmar script. MICR (Myanmar Intelligent Character Recognition) is a technique based on ICR. High speed recognition rates can be gained by using MICR. It was trained and experienced with successfully in both typeface and handwritten characters. It is used to recognize effectively hand-printed characters. But, it also has broken, overlapped and noisy characters. Hybrid approach can successfully overcome these problems. In this paper, Hybrid approach is used the result of MICR and back-propagation neural network (BPNN). BPNN is a good recognition engines. It has the ability to generalize by making decisions about imprecise input data. It also offer solutions to a variety of classification problems such as speech, character and signal recognition. The arrangement of this paper is as follows: Section (2) express the Myanmar Language, Section (3) explain Myanmar character research field and MICR, Section (4) show the composition of the proposed system, Section (5) present the new contribution of Hybrid approach, Section 6) show output. Experimental results and conclusion are in Section (7) and (8), respectively. 2. Myanmar Language The Myanmar language is the official language and is more than one thousand years old. Myanmar script is considered a complex script by software developers, as it

43 23 originated from Indic scripts like Thai or Khmer. The Myanmar (formerly known as Burmese) script developed from the Mon script, which was adapted from a southern Indian script during the 8th century. The earliest known inscriptions in the Burmese script date from the 11th century. Myanmar alphabet consists of 33 consonants, 12 vowels, 4 medials, 10 digits and a lot of Pali character as shown in Figure 1. but practical research is only a few works in research field. Because, the problem of Myanmar characters recognition is more difficult than English languages in respects including the similarity of characters, absence space between each word, etc. So, various character recognitions method not enough complete recognize Myanmar character. They are still in research field, not complete work. MICR is an interesting algorithm to recognize Myanmar characters that has been developed recently in Myanmar. 3.1 MICR Figure 1. A set of Myanmar alphabet In Myanmar (Burmese) writing system: syllabic alphabet - each letter has an inherent vowel. Other vowels are indicated using separate letters or diacritics which appear above, below, in front of, after or around the consonant. The rounded appearance of letters is a result of the use of palm leaves as the traditional writing material. Straight lines would have torn the leaves. The Burmese name for the script is 'round script', is written from left to right, as shown in Figure 2. C K T V P MICR (Myanmar Intelligent Character Recognition) system is one category of ICR (Intelligent Character Recognition) methods and it can recognize not only online characters but also off-line too. But it is more suitable for noise free images and isolated characters. It used statistical and semantic approach to collect information. That information includes the data of width and height ratio, horizontal and vertical black stroke count, number of loops, end point, open direction, histogram values and character type, etc. After gathering this required information for each character, we put them on the properties array to record them. Properties of each character are compared with Pre- Defined Database: Basic characters (B-database), Extended characters (E-database), Medials (M-database). When the incoming character matches with the database, the voting system is used to make the final decision of the image on that information. If the incoming character is equivalent to the predefined database, the voting system produces the relative code number for that character. This code number is stored in the code buffer. Otherwise, reject message is generated, as shown in Figure3. No. of char<=n No Yes Statistical and Semantic approach M C Consonants P Punctuations K Killer T Tone V Vowels M Medials Figure 2. Terms of Character s samples C No Compare Voting B[0]B[1].B[n]. E[0]E[1].E[n]. M[0]M[1]...M[n]. Basic Extended Medials 3. Myanmar Character Research Area The interests in Myanmar handwritten characters recognition research have grown over the past few years Reject Yes Code Figure 3. System schematic diagram Information Figure 3. MICR System schematic diagram

44 Applications of MICR MICR has been successfully applied in a lot of application such as: Speed limited road signs recognition Car license plate reader Recognition of Myanmar basic characters and compound words(vowels) Online Myanmar medial hand-printed characters into machine editable text On-line Handwritten Myanmar Pali Character recognition Handwritten English Characters to Machine editable text by applying MICR Converting Myanmar Portable Document format to machine editable text with format using MICR Voice production of Handwritten Myanmar Compound Words Enhancing the Myanmar Pali Recognition based on MVM (Myanmar Voice Mixer) 4. System Composition This section describes the simple technique involved in our proposed online handwriting recognition system. This is a writer-independent system based on the Hybrid approach. In this system includes four stages: Data acquisition, Pre-processing, Hybrid approach and Output, as shown in Figure 4. Data Acquisition Pre-processing Hybrid Method Statistical & Semantic information of MICR Back-propagation Neural Network 4.1 Data Acquisition Two different types of data input method: online and offline. In online data acquisition, data inputs are stored as images that are concurrently written by the users through Tablet, hand-held PDA devices, etc. In offline data acquisition, data inputs are stored as images that are captured by scanner., the proposed system can handle on only online data input by users. Isolated characters are needed to process the image. 4.2 Pre-processing Various preprocessing operation are: Gray Scale Converting, Noise Filtering, Binarization, Extraction, Resizing and Normalization. Firstly, convert the incoming original image into gray level image and then filtering the noise of the image result from gray scale conversion of image. If conversion of a grayscale image into a binary image, the system extract row and column for each character. In normalization, it is performed on the digitized image to enhance the quality of the image. Then, labeling scheme is used in this system for the one character lonely. 5. Hybrid Method The system used to combine the MICR and back propagation neural network. Each of these algorithms has its own specific strengths and weakness. The idea of Hybrid method in order to compensate their individual weakness and to prevent their individual strength has been widely used in Myanmar character recognition field. In this system, back-propagation neural network uses the features of statistical and semantic information of MICR as input nodes. So, training time is very quickly and received the high accuracy rate. 5.1 Statistical and Semantic Information of MICR After the pre-processing stage, feature information of each character is extracted by using MICR. MICR used statistical and semantic approach to collect information such as width and height ratio, horizontal and vertical black stroke count, number of loops, histogram values and character type, etc, shown in Figure 5. Change Unicode / ASCII code Figure 4. Proposed system design Output A statistical approach looks for a typical spatial distribution of the pixel values that characterize each character. It is searching for the statistical characteristics of various characters. These characteristics could be very simple, like the ratio of black pixels to white pixels, width and height ratio, histogram, etc. Some of handwritten characters indeed consist of pixels. Statistical methods ignore is that the pixels also form lines and contours. A

45 25 semantic approach recognizes the way in which the contours of the characters are reflected in the pixels that represent them and try to find out typical characteristics for each character. Semantic data: black stroke count, loop, open, end point, etc. statistical and semantic information of MICR has been used the inputs of neural network to save the training time. Width and Height Ratio Loops 2rows 1col 1row 1col 1row 1col 1row 2cols Input layer Hidden layer Output layer Figure 6. Network architecture No loop Loop count-2 Loop count Back-propagation Algorithm Horizontal Black Stroke Count Vertical Black Stroke Count Open Direction P5 P4 P3 P2 P1 First, the training sample is fed to the input layer of the network. For unit j in the input layer, its output is equal to its input, that is, Oj I for input unit j. The net input to j each unit in the hidden and output layers is computed as a linear combination of its inputs. A unit j in a hidden or output layer, the net input, I j, to unit j is I j wijo (1) i where w ij is the weight of the connection from unit i in the previous layer to unit j; O i is the output of unit i from the previous layer. The net input I j to unit j, theno j, the output of unit j, is computed as No open Open at P3 Open at P1 & P2 Endpoints P6 P7 P8 O = (2) j 1 1+e -Ij For the output layer, the error value is: j = O j (1- O j )(Tj - O j) (3) And for hidden layer Figure 5. Some statistical and semantic information 5.2 Back-propagation Neural Network In this system, feed-forward neural network and backpropagation learning algorithm is used. The recognition performance of the Back propagation network will highly depend on the structure of the network and training algorithm. It is consists of three layers forward structure that has hidden layer between input layer and output layer interconnected by links that contains weights. Figure 6, shows the architecture of network. Its input has two forms: features extraction of image and pixels of image. Training time will be very long if pixels are used as an input for neural network. In this system, where j = Oj (1-Oj ) kwjk k (4) wjk is the weight of the connection from unit j to a unit k in the next higher layer, and k is the error of unit k. Weights are updated by the following equations, where w ij is the change in weight w ij : w O (5) ij j i wij wij w (6) ij The variable β is a constant learning rate. The parameters used in the back-propagation neural network experiments are listed in Table1.

46 26 Table 1. Parameters used for back propagation neural network Parameter Values Handwritten Character Input Layer neurons 165 Hidden Layer neurons 100 Output Layer neurons 6 Number of Epochs Performance Function Mse (Mean square error) Back-propagation learning Rate 0.1 Momentum Term 0.9 Minimum Error Exist in the Network Initial weights and biased term values 0.01 Randomly Generated Values Between 0 and 1 Editable Text 5.4 Change Unicode/ASCII Code In this step, the code numbers of characters are changed into their relative native code (Unicode or ASCII) for output. Some Myanmar character and their relative code sequences are shown below: Table 2. Sample words and their native code 7. Experimental Result In order to check the accuracy of the individual recognition method, handwriting samples were collected from various people. In this paper, the system was trained and tested over 1000 sample. Table 3. Recognition rate for MICR No of samples Recognition accuracy rate for noisy image Recognition accuracy rate for noise free image % 94% % 92.7% 6. Output After that, the recognized combined words are produced as output. This output can be shown in the Microsoft Word file as the editable text format. To connect Microsoft Word, rich text format (rtf) function is used. The output result for before recognition and after recognition output file is provided as example: % 90% % 87.2% % 85% Shown in Table (3) is the recognition accuracy rate of MICR recognizer. This table shows noisy and noise free image of handwritten Myanmar characters and digits.

47 27 Table 4. Recognition rate for Back-propagation Neural Network No of samples Table (4) shows the recognition accuracy rate of backpropagation neural network recognizer. Pixels of image are used input node for neural network. Table 5. Recognition rate for Hybrid The hybrid method was used to combine the MICR recognizer and back-propagation neural network to create the high accuracy rate. Features of image are used input node for this method. So, processing time is very quickly. 8. Conclusion Using Pixel Input Recognition accuracy rate for noisy image Recognition accuracy rate for noise free image % 91% % 88.75% % 85% % 83.41% % 81% No of samples Using feature Input Recognition accuracy rate for noisy image Recognition accuracy rate for noise free image % 98.89% % 97% % 95.85% % 94.50% % 93% Table 6. Average processing time for recognizers Kinds of recognizer No of samples Average processing time MICR recognizer seconds Back-propagation recognizer seconds Hybrid recognizer seconds The paper has presented a new contribution of handwritten recognition using a robust combination of hybrid approach through MICR (Myanmar Intelligent Character Recognition) and back-propagation neural network. This paper has compared the forecasting accuracies of MICR recognizer, back-propagation recognizer and hybrid method. Hybrid method can recognize not only similar characters in different language but also different handwritten styles. In this system, it can recognize only Myanmar handwritten characters. Acknowledgments The authors would like to thank all participants in MICR (Myanmar Intelligent Character Recognition) research field from the University of Computer Studies in Yangon. We would like to thank the reviewers and editors of IJCSI. Finally, we appreciate all the readers who spend their precious time to read this paper. References [1] E.Phyu, Z.C.Aye, E.P.Khaing, Y.Thein and M.M.Sein, Recognition of Myanmar Handwritten Compound Words based on MICR, the 29th Asian Conference on Remote Sensing (ACRS), Colombo, Sri Lanka, 2008 [2] Hamza, Ali A, Back Propagation Neural Network Arabic Characters Classification Module Utilizing Microsoft Word, Journal of Computer Science 4 (9): , 2008 [3] Zin Lynn Phyu, Yin Mon Aung, Ei Phyo Min, Yadana Thein, Hybrid of ICR/OCR technology through MICR and Neural Network, 30th Asian Conference on Remote Sensing(ACRS), China, 2009 [4] Ei Phyo Min, Yadana Thein, Applying a Hybrid OCR/ICR technology through Neural Network and MICR to Handwritten Character Recognition, University of Computer Studies, Yangon [5] Srinivasa Kumar Devireddy, Settipalli Apparao, Hand Written Character Recognition using Back Propagation Network,Journal of Theoretical and Applied Information Technology [6] Sang Sung Park, Won Gyo Jung, Young Geun Shin, Dong-Sik Jang, Optical Character Recognition System Using BP Algorithm, IJCSNS International Journal of Computer Science and Network Security, VOL.8 No.12, December 2008 [7] Tun Thura Thet; Jin-Cheon Na, Wanna Ko Ko, Word Segmentation of Myanmar Language, Journal of Information Science JIS, 2nd October, 2007 [8] Zaw HTUT (Mr.), Features of Myanmar Language Document Styles, Executive Committee Member, MCSA Myanmar Computer Federation (MCF) Dr. Yadana Thein I received M.Sc (Master Computer Science) degree in 1996 and PhD (I.T) degree in I am now associated professor of U.C.S.Y (University of Computer Studies, Yangon). I have written about 25 papers altogether. About 10 of them are local papers and 15 are foreign papers. My first paper is Recognition of Myanmar Handwritten Digits and Characters for ICCA conference in My research interests include Image Processing, Neural Network and MICR (Myanmar Intelligent Character Recognition) field. San Su Su Yee I am a Master Thesis Student. I received B.C.Tech degree in 2008 and B.C.Tech (Hons.) degree in I got one distinctions (English) in Master course work exam.

48 28 Ontology Based Agent Communication in Resource Allocation and Monitoring Manish Arora 1 and M. Syamala Devi 2 1 DOEACC Society Chandigarh Centre, Chandigarh, India 2 Department of Computer Science and Application, Panjab University Chandigarh, India Abstract The aim of ontology is to share information between sending and receiving agents of Multi Agent System (MAS). It provides standard vocabulary and terms for knowledge sharing and is designed to share information conveniently and understandably. Agent based application requires complex interaction among agents. This complexity is due to agent-agent and agent-user communication. It is required to use ontology in agent based application of resource allocation and monitoring. The purpose of Resource Allocation and Monitoring System is to make the procedures involved in allocating fund resources to competing fund seekers transparent so that deserving candidates get funds. Proactive and goal directed behaviour of agents make the system transparent and intelligent. This paper presents ontology designed and implemented for the purpose of communication among agents of Multi Agent System for Resource Allocation and Monitoring (MASRAM). FIPA (Foundation for Intelligent Physical Agents) compliant software JADE (Java Agent Development) is used to implement ontology. Keywords: Multi Agent System, Resource Allocation, ACL, FIPA, JADE, Ontology. 1. Introduction Ontology is used to represent knowledge that is shared between different entities. It provides terms and vocabulary used to represent knowledge so that both sender and receiver can understand. Ontology is widely used in many areas like MAS and Biomedical Informatics to share knowledge. The study of MAS focuses on systems in which many intelligent agents interact with each other. The agents are considered to be autonomous software or hardware entities, such as software programs. Their interactions can be either cooperative or selfish. Agent acts on the behalf of users / other agents with different goals and motivations. Agents require ability to cooperate, coordinate and negotiate with each other to complete task successfully [1]. Agents work independently but share information. Ontology helps designers of agent based systems to make information understandable between agents. Ontology is applied in one such agent based application, resource allocation and monitoring. Resource allocation problem occurs when fixed and limited resources are allocated to competing fund seekers to execute their projects. These resources may be of different types like work force, machine timings, raw material and funds. Fund seekers can submit their project proposals to avail grant to allocating agencies. Projects can be of different nature like R & D projects and social oriented schemes. On receiving the project proposals from fund seekers, fund allocating agencies evaluate proposals technically as well as financially. After the submission of fund request, committee on the behalf of funding agency evaluates proposals. In some cases, fund seekers are asked to present the project proposal. Based on the recommendation of committee, funds are allocated by considering both quantifiable and non-quantifiable factors. Considering above facts, an integrated decision making system Multi Agent System for Resource Allocation and Monitoring (MASRAM) is designed and agents of system require information sharing. This paper describes the ontology based communication between agents of MASRAM. The designed ontology is implemented in JADE. The paper has been organized into different sections. Second section reviews the related research, third describes the model, fourth details construction of ontology and fifth shows how ontology is implemented in JADE.

49 29 2. Review of Related Research A fundamental characteristic of MAS is communication among different agents operating in the system. Agents exchange information in order to achieve their goals. Message follows Agent Communication Language (ACL) standard which allows encoding/decoding of actual message. The structure of the message is set of terms written in FIPA-ACL like message content, message parameters, encoded message and transports layer information. Contents of messages are written in content language such as FIPA-SL (Semantic Language) and FIPA-KIF (Knowledge Interchange Format). FIPA-ACL is based on Speech Act Theory which means that message represents action or communication act, also known as performative act. Other commonly used communication acts are inform, request, agree and refuse [2]. When agents in MAS communicate with each other, message is sent and main component of message is content slot. According to FIPA specification, value of this slot could be either string or raw sequence of bytes. In real world application, agent needs to send complex information to the receiving agent like list of agencies providing funds. In such scenario, a well defined syntax of the message content is adopted so that both sender and receiver agents can understand and share information. The concept is called Content Language. Two kinds of languages are used SL and Leap. Our research work is based on FIPA-SL content language. This language is used to define concept and symbols used in content of the message and is known as Ontology. Ontology is agreement about shared conceptualization that includes framework for modelling application specific contents for communication among agents [3]. The aim is to clarify meaning of the message for exchange. Jenyl Mumpower and Thomas A Darling [4] have discussed three procedures that can be used to resolve Resource Allocation problem. In Incremental Appropriation, resource allocation begins with no allocation and then allocates small resources. The process is repeated until resources are exhausted. In the second procedure, multiple negotiators give different concessions. Resources are moved from one point to another and utility function is checked. In the third procedure, different negotiators assign different weights to different programmes. Quantification of non-quantitative factors is important to make decision of allocation [5]. The non-quantifiable factors can be measured through fuzzy comprehensive measurement method. Since non-quantifiable factors are measured by human whose knowledge and experiences may not be exact or complete. The probabilistic tools are used to deal with such data. This approach is also used to rank employees performance using both quantitative and non-quantitative measures. Monitoring is very important factor to know the utilization of the funds, benefits gained from funding and giving further financial help. Various agent oriented tools are available to develop intelligent agents that include basic services like communication act. One of them is JADE. It is FIPA compliant tool and same is used to develop agents defined in MASRAM problem. JADE simplifies the implementation of MAS through middle tier. Message in JADE can be passed through string data type, Java object or ontology. The focus of research is to construct and implement ontology comprising complex information in JADE [6]. From review, it has been found that MAS systems are widely used in resource allocation problems, such as transportation, scheduling, production planning and system resources in which ontology plays important role in communication between agents [7, 8]. 3. MASRAM Model Three agents have been designed for MASRAM problem. At the abstract level these agents are:- Coordinator Agent Fund Seeker Agent Fund Allocator and Monitor Agent 3.1 Coordinator Agent Coordinator Agent interacts with three types of users of MASRAM i.e. Fund Seeker user, Fund Allocator user and Reviewer user. Fund Seeker user seeks funds, Fund Allocator user allocates funds and monitors the utilization while Review user reviews the proposal. Coordinator Agent forwards requests received from Fund Seeker user to Fund Seeker Agent. Coordinator Agent also forwards requests received from Fund Allocator user and Reviewer user to Fund Allocator and Monitor Agent. To summarize, this agent interacts with following agents/users. Fund Seeker User Fund Allocator User Reviewer User Fund Seeker Agent

50 30 Fund Allocator and Monitoring Agent Fund Seeker User (s) Fund User (s) Allocator Reviewer User(s) 3.2 Fund Seeker Agent Fund Seeker Agent receives all the requests received from Coordinator Agent and act accordingly. This agent interacts with Coordinator Agent only. 3.3 Fund Allocator and Monitor Agent MASRAM Coordinator Agent Fund Allocator and Monitor Agent in turn evaluates proposal, assigns weights and allocates suitable funds based on allocation procedure. Fund Allocator and Monitor Agent (FAMA) interacts with Coordinator Agent only. Fund Agent Seeker Fund and Agent Allocator Monitor Agents and overview of the interaction among them have been shown is figure 1. MASRAM Database 4. Design of Ontology Ontology for multi agent based application described above is designed and implemented in JADE. After the agents have been defined, next step considered was to design ontology that includes set of concepts and symbols. Table 1 describes the messages passed to Fund Seeker Agent from Coordinator Agent along with type of communication act e.g. source_type is message of communication act request. Coordinator Agent uses this message to request Fund Seeker Agent to provide list of funding agencies based on particular type of project. Similarly fund_utilization message is used to inform Fund Seeker Agent the status of utilization. Other content messages are also designed based on responsibilities of agent in similar fashion. Ontology was also designed for Coordinator and Fund Allocator and Monitor agent to communicate with each other. Table 2 describes one such content message, fund_category that is passed to Fund Allocator and Monitor Agent by Coordinator Agent. Coordinator Agent informs various fund categories out of which funds can be provided. Table 3 and 4 show some responses given by Fund Seeker Agent and Fund Allocator and Monitor Agent after receiving requests from Coordinator Agent. Fig 1. Multi Agent System Resource Allocation and Monitoring Model. Table 1: Message passing (coordinator-fund seeker) Sender Agent: Coordinator Receiver Agent: Fund Seeker Content Messgae Act Source_type Request Fund_utilization Inform Table 2: Message passing (Coordinator-Fund Allocator) Sender Agent: Coordinator Receiver Agent: Fund Allocator and Monitor Content Message Act Fund_category Inform Table 3: Message passing (Fund Seeker-Coordinator) Sender Agent: Fund Seeker Receiver Agent: Coordinator Content Message Act Available_source Inform Proposal_id Inform

51 31 Table 4: Message passing (Fund Allocator and Monitor-Coordinator) Sender Agent: Fund Allocator and Monitor Receiver Agent: Coordinator Content Message Act Allotted_fund_categor y Inform Review_proposal Request One of the content messages described above is detailed in table 5. Table 5: Ontology Available_source (Inform : sender (agent-identifier :name FSA@manish:1099/JADE) : receiver (agent-identifier :name COA@manish:1099/JADE) : ontology ra-ontology : language fipa-sl : content ( Available_source: (Sequence ( ( nature_of_project_id : 1 Nature_of_project_desc: Information and Technology allocator_id: 2001 allocator_desc: Department of Information Technology allocator_address: New Delhi criteria: (sequence ( 1, Number of Technical Staff Members,5) (2, Projects Handled, 5) ) Fund_available : (sequence ( )) ) ) ) ) The message shows that Fund Seeker agent named FSA wanted to send list of allocators who are funding project related with Information and Technology. It sends criteria set by allocator. Agent also sends availability of funds. Ontology mentioned here is application specific ontology (ra-ontology). On the similar lines remaining messages are defined. 5. Implementation in JADE To implement agents and ontology, JADE tool is used [9]. Following steps describe implementation procedure. 5.1 Developing Ontology Ontology in JADE is instance of jade.content.onto.ontology class. Ontology is collection of schemas. For MASRAM problem, conceptschema is used in Java code that extends Ontology class of JADE. Table 6 shows the code written in Java. Name of the ontology is defined as ra-ontology. Five schemas have been defined. LOGIN and PWD schema are used for user verification, PROJ_NATURE is used to pass type of the project e.g. Education and Research, PROJ-SOURCE is used to send list of the fund allocators along with their criteria and availability. Lastly PROJ_LIST is used to send list of project types. First three schemas are of primitive in nature while remaining two are of the aggregate schema in which list is passed as message. ObjectSchema.UNLIMITED means any number of rows can be sent. 5.2 Developing Java Classes While implementing ontology in JADE, each schema is associated with java class that implements AgentAction class of JADE. Each Schema can have one Java class for each schema or can have single Java class for multiple schemas. In MASRAM case, one single java class file is used. Each schema has two public declared methods. One method is setxxxx and other is getxxxx where XXXX is name of schema. These two methods are used to set and retrieve the values respectively as briefed in table 7. Table 6: Defined ontology listing // import statements public class RAOntology extends Ontology { public static final String ONTOLOGY_NAME = "ra-ontology"; public static final String LOGIN = "Login_id"; public static final String PWD = "Login_pwd"; public static final String PROJ_SOURCE = "Proj_source"; public static Ontology getinstance () {return theinstance; } private RAOntology() { super(ontology_name,basicontology.getinstance()); try { add(new ConceptSchema(LOGIN),Login.class); ConceptSchema cs = (ConceptSchema)getSchema(LOGIN); cs.add(login,(primitiveschema)getschema(basicontology.strin G)); cs.add(pwd,(primitiveschema)getschema(basicontology.string), ObjectSchema.OPTIONAL); cs.add(proj_source,(aggregateschema)getschema(basicontolog y.sequence),0,objectschema.unlimited); Passing Message Using Ontology Third step is to call defined ontology in Java agent program to fill and send message. Table 8 shows the important lines of code of setting message. The java program imports ontology defined earlier along with other packages. RAOnotlogy.getInstance() in code sets user defined ontology. Table 7: Listing of Java class

52 32 // import statement public class Login implements AgentAction { String login_id; String login_pwd; List proj_source; public void setproj_list(list l) { proj_list=l; } public List getproj_list() { return proj_list; } Table 8: Listing of setting message Login lg = new Login(); Types ty = new Types(); ACLMessage msg = new ACLMessage(ACLMessage.REQUEST); getcontentmanager().registerlanguage(codec); getcontentmanager().registerontology(ontology); List l = new ArrayList(); msg.setlanguage(codec.getname()); msg.addreceiver(new AID(board.getReceiver(), AID.ISLOCALNAME) ); msg.setlanguage(codec.getname()); msg.setontology(ontology.getname()); msg.setontology(ontology.getname()); action.setactor(new AID(board.getReceiver())); action.setaction(lg); this.getcontentmanager().fillcontent(msg,action); send(msg);.. 6. Conclusion This paper detailed the way agents interact with each other through ontology. During construction of ontology, FIPA specifications were followed and implemented in FIPA compliant agent development framework, JADE. Ontology was developed in java classes by importing JADE packages. A three layer approach is used. In first layer, java class file defines terms (objects) to be used. In second layer, Ontology and schemas are defined and in third layer, message is filled. This method of ontology found suitable to exchange complex information like multiple records and made them understandable. 7. Scope for Future Work Future plan includes implementing fund allocation algorithm in JADE so that agent can allocate funds among competing fund seekers. The complete developed MASRAM will be tested to validate the work done. winter2006/slides/04-masdef-handout.pdf, accessed on April 11, [2] History of FIPA, available at FIPA.htm accessed on Sep 11, 2010 [3] Prashant M, Integrating Ontologies into Multi-Agent Systems Engineering (MaSE) for University Teaching Environment, Journal of Emerging Technologies in Web Intelligence, Vol 2, No 1, 2010, pp [4] J. L. Mumpower, and T. A. Darling, Modeling Resource Allocation Negotiations, in IEEE Twenty Fourth Annual Hawaii International Conference, 1991, Vol. 3, pp [5] J. Cheng, H. B., and Ziping Li, Quantification of Non Quantitative Indicators of Performance Measurement, in 4 th International Conference on Wireless Communications, Networking and Mobile Computing WiCOM '08, 2008, pp [6] JADE Administration s Guide, available at accessed on June 30, [7] V. Gorodetski, Oleg Karsaev, and Victor Konushy, Mulit Agent System for Resource Allocation and Scheduling, in 3rd International Workshop of Central and East European conference on Multi Agent System, Prague, Czech Republic, 2003, Vol. 2691/2003, pp [8] Anthony Chavez, Alexandros Moukas and Pattie Maes, Challenger: A Multi Agent System for Distributed Resource Allocation, in First International Conference on Autonomous Agents (Agent97): Marina Del Ray, California: ACM Press, 1997, pp [9] Fabio Bellifemine et el, JADE A FIPA Compliant agent framework, available at /viewdoc/download?doi= &rep=rep1&type=pdf, accessed on July 20, Manish Arora holds MCA (GNDU, Amritsar), MBA in Operation Management (IGNOU) and C Level (M. Tech) from DOEACC Society, New Delhi. He has nearly two decades experience in teaching, software development and consultancy. Presently, he is working as Principal Systems Analyst in DOEACC Society, Chandigarh Centre and managing different government projects. He has published 3 papers on multi agent technologies in international journals and conferences. Dr. M. Syamala Devi is a professor in the Department of Computer Science & Applications, Panjab University, Chandigarh (India). She received her Ph.D degree in Computer Science and Systems Engineering from Andhra University, Visakhapatnam and M. E. in Computer Science & Engineering, from NIT, Allahabad. She had also served ISRO, Sriharikota, and NITTTR, Chandigarh. Her areas of expertise include algorithms Design and analysis, Image Processing, Distributed AI. She has about fifty national and international research papers to her credit. References [1] Jarg Denzinger, Multi Agent Systems, Department of Computer Science, University of Calgary, Canada, available at

53 33 A Novel DSS Framework for E-government A.M. Riad 1, Hazem M. El-Bakry 2 and Gamal H. El-Adl 3 1 Information System Department, Mansoura University, Mansoura, Egypt 2 Information System Department, Mansoura University, Mansoura, Egypt 3 Information System Department, Mansoura University, Mansoura, Egypt Abstract There are various frameworks for decision support system (DSS) that have been formulated. Such frameworks describe the characteristics of DSS. In this paper a proposed effectiveness DSS framework for e-government is presented. This is done by integrating DSS components to support managers and decision makers in e-government. The structure of the proposed DSS framework is discussed. It is expected that the performance of proposed framework will be more effective than existing e- government systems. Keywords: Decision Support Systems (DSS), Frameworks, electronic-government (e-government) 1. Introduction A variety of DSS have been postulated for describing the characteristics of DSS. These frameworks are helpful in organizing and identifying the relationships of DSS. The identification of DSS applications is important in planning organization strategy for the deployment of information technology. DSS is defined as the use of computer to: (i) Assist managers with their decision process in semistructure tasks; (ii) To support, rather than replace managerial judgment, and (iii) To improve the effectiveness of decision making rather than its efficiency [1,3]. While definitions of e-government by various sources may vary widely, there is a common theme. E-government involves using information technology, and especially the Internet, to improve the delivery of government services to citizens, businesses, and other government agencies to interact and receive services from the federal, state or local governments twenty four hours a day, seven days a week [7,8]. E-government involves the use of information and communication technologies (ICTs) to support government operations and provide government services [5]. There is a relation between E-government and DSS where E-government encourages citizen participation in the decision- making process and making government more accountable, transparent and effective [3]. The problem focuses on where is the decision support system into the e-government components/ layers and how to utilize the useful of DSS into e-government. No explicit e- government framework includes DSS into its components. So the proposed framework used to solve this problem. In this work, a novel proposed DSS framework for e- government is presented by integrating its components into the e-government framework layers. The aim is to utilize its components to help decision-makers within the e-government. This paper is organized as follows. Section 2 presents an overview about DSS and its framework. Section 3 reviews e-government in general while section 4 describes its partnerships. Section 5 shows the e-government layers. The discussion of the proposed framework layers is given in section 6. Finally the conclusion is presented in section Decision Support System an Overview DSS is an interactive, flexible, and adaptable computer based information system. It has been developed for supporting the solution of a non-structured management problem and improving decision making. It utilizes data,

54 34 provides easy user interface, and allows for the decision maker s own insights [1,2]. DSS provides support in semi-structured and unstructured situations, includes human judgment and computerized information. DSS supports various managerial levels as in Fig.1 [3]. Fig. 1 Decisions in Management Levels [3]. DSS components are: (i) database management subsystem (DBMS), (ii) model base management subsystem (MBMS), (iii) knowledge-based (Management) Subsystem, and (iv) User interface subsystem (Dialogue). Fig.2 [4] shows the components of the DSS. DSS is a problemsolving tool and is frequently used to address ad hoc and unexpected problems. DSS is one of information system types, so any computerize application indeed is a DSS as e-government services where available on the Internet daily 24 hours. Database DBMS MBMS Dialogue User Interface System. User Model Base Fig.2 Components of the DSS [4]. 3. Definitions of E-Government To understand E-government, it must understand administrative development and reform on government in general. During two decades, administrative reform and development have experienced total quality management (TQM) in1980s, and reengineering and reinventing government in 1990s [6]. Government reinvention make us realized that government is actually a dynamic mixture of goals, structures and functions. E-government initiatives are complex change efforts intended to use new and emerging technologies to support a transformation in the operation and effectiveness of government derived from government reinvention. New challenge of public administration in 2000s or 21st century is to create an E- government. E-government is defined as: government activities that take place over electronic communications among all levels of government, citizens, and the business community, including: acquiring and providing products and services; placing and receiving orders; providing and obtaining information; and completing financial transactions [8]. E-government is the continuous optimization of service delivery, constituency participation and governance by transforming internal and external relationships through technology, the Internet and new media. This includes government to citizen, government to employee, government to business, and government to government. 4. Types of E-Government Partnerships Summarized from our research on e-government, normally, government identifies and drives implementation of eight types of E-government which can bring significant benefits to the Government, citizens, business, employees and other nonprofit organizations and political and social organizations. Types of E-Government can be classified into 8 categories, are as follows: (i) Government-to-Citizen (G2C): Provide the momentum to put public services online, in particular through the electronic service delivery for offering information and communications; (ii) Citizen-to-Government (C2G): Provide the momentum to put public services online, in particular through the electronic service delivery for exchange of information and communication; (iii) Government-to-Business (G2B): Actively drive E- transactions initiatives such as e-procurement and the development of an electronic marketplace for government purchases; and carry out Government procurement tenders through electronic means for exchange of information and commodities; (iv) Business -to-government (B2G): Actively drive E- transactions initiatives such as e-procurement and the

55 35 development of an electronic marketplace for government purchases; and carry out government procurement tenders through electronic means for sale of goods and services; (v) Government-to-Employee (G2E): Embark on initiatives that will facilitate the management of the civil service and internal communication with governmental employees in order to make e-career applications and processing system paperless in E-office; (vi) Governmentto-Government (G2G): Provide the Government's departments or agencies cooperation and communication online base on mega database of government to have an impact on efficiency and effectiveness. It also includes internal exchange of information and commodities; (vii) Government-to-Nonprofit (G2N): Government provides information and communication to nonprofit organizations, political parties and social organizations, Legislature, etc, and (viii) Nonprofit-to-Government (N2G): Exchange of information and communication between government and nonprofit organizations, political parties and social organizations, Legislature, etc. from the above categories of E-government, we can sum up that E- Government initiatives should focus on five consumer-togovernment relationships: Citizen-to-Government, Business-to-Government, Government-to-Nonprofit, Government-to-Government and Government-to- Employee. First, Citizen-to-Government refers to the direct consumption of public services by the individual consumer for personal use. These services include licensing and permitting for hunting, fishing, and driving privileges. This will not only include the payment of taxes, fines, and fees to state and local governments, but also the payment of refunds to taxpayers. Second, the Business-to- Government relationship model refers to those services consumed by entrepreneurs, businesses, and corporations, for a commercial purpose (profit or nonprofit). These include filing statements of incorporation, obtaining business licenses, assistance with site locations, and obtaining workforce information. Finally, Government-to- Nonprofit, Government-to- Government and Governmentto-Employee refer to the coordination of both inter- and intra- agency cooperation and employees to improve services inside or outside governments. This includes travel requests, purchasing requisitions, payroll processing, intergovernmental fund transfers, and position applications, etc [8, 9]. 5. E-Government Layers Fig.3 [7] gives an overview of the system architecture. The security layer is based on a public key infrastructure1 (PKI) that provides a nation wide electronic identity. The interface layer comprises all components needed to interact either with human or non-human users (other systems). The actual business logic (e.g. logic for document and form management, signature verification, plausibility checks) is contained in the function layer. The transaction layer forms an abstraction of all possible underlying backend-system [6, 7, and 10]. It provides a uniform application programming interface (API) that is implemented by different backend-system adapters. Every service is, depending on the underlying business process, bound to a specific backend-system. The transaction manager automatically selects the appropriate adapter, thus all systems can be treated uniformly. The range of supported backend-systems ranges from ordinary to sophisticated workflow-management solutions. To further support the scalability of the system and to minimize the impact of changes on the entire system, individual components had to be decoupled as far as possible. This was achieved by using technologies and data formats that had originally been introduced to support the communication between different e-government systems also for the internal data representation. The platform provides the following basic functionality: (i) Start of new procedures by the submission of electronic application forms (including attachments). Depending on the specific use case forms might have to be digitally signed; (ii) Electronic notification service concerning the progress of the underlying business process; (iii) Electronic payment; (iv) Overview of the current state of all procedures (requires login using the PKI card), and (v) Electronic delivery. Fig.3 E-government layers [7]. The question now is "Where is DSS in e-government framework?". So if we make a combination between DSS framework and e-government framework, the result will be integrated platform that satisfies the managers requirements. Moreover, the decision making process will satisfy the user requirements. The proposed framework of e-government is shown in Fig. 4.

56 36 6. Structure of the Proposed Framework The Structure of the proposed system can be summarized as: (i) Access channels: It means the way a user visits E- government portal. Users can use different kinds of terminal devices, and enjoy the personalized E- government services permitted any time and any place. This can be done by using different access channels. The terminal devices that users can utilize include: PCs, portable computers, cell-phones, common telephones, PDAs. The access channels mainly include: Internet, government private network, information kiosk, telephone, TV, digital TV, and mobile communication, etc; (ii) E-government portal: is the uniform entry-point for the entire E-government system. It has an interactive main entrance for all kinds of users to find the services they need; (iii) Applications layer: The particular E- government systems are constructed by functional departments mainly include vertical systems of functional departments, cross-departmental systems and comprehensive decision support systems for decisionmakers within the e-government that imbed DSS. Components of DSS are based on the data that stored in the DB and also data models that generated from the manager experience in taking the decisions; (iv) Unified application support platform: Between the application and resource layers, the unified support platform plays a significant role in forming a connection in the E- government overall technical framework. It is an open infrastructure independent from the networks and its applications; (v) Information resources of government affairs: Built on top of the network infrastructure, the information resources layer offers various kinds of information resources of government affairs in its upper layer, the unified application support platform layer. The government information resources are mainly composed of shared information resources, catalogue resources and the internal thematic information resources of each department, where shared information resources is combined with basic information resources, shared business information resources and comprehensive information resources while catalogue resources are made up of data catalogue, service catalogue and user directory, and (vi) Government network layer: It facilities an important infrastructure level to support E-government which includes wired as well as wireless private networks at municipal, district and county levels, vertical departmental private networks constructed by departments of different functions according to departmental characteristics and needs, and various public website resources. Table (1) shows a comparison between the previous e-government framework and our proposed framework. During many different interviews for e- government offices; the managers asked for clear DSS tools in e-government. This is prove that the performance of our proposed framework is more effective than existing e-government systems. The majors in table (1) clarify that the proposed framework is very useful for decision makers in e-government. Top Managers in e-government need accurate decisions quickly. The proposed framework helps top manager to do all supervision tasks in e-government efficiently. Table 1: Comparison between previous and our proposed e-government framework. Majors S E C U R I T Y Previous Framework Proposed framework Top management Low High 7. Conclusions Access Channels Portal Layer Applications Layer Unified App. Support Platform Information Resources Layer Network Layer Fig.4 The proposed integrated DSS with e-government framework. Accurate decisions Moderate High Provision of decision Low assurance High assurance DSS Supervision in e-gov Less effective More effective We have shown that the movement to e-government is very important for government to interact with people and business transactions. E-government offers a huge potential to find innovative way to reach the satisfaction of people. Progress of new technologies allows electronic services to be applied in e-government. So DSS must be integrated with e-government managerial levels. We have clarified that DSS is a very helpful tool for all e- government partnerships. It was known that DSS

57 37 frameworks are not included into e-government framework layers. Here, it has been proved that the application layer should be combined with DSS framework to make use of its benefits for top mangers. References [1] Keen, P. and Scott-Morton, M. "Decision Support Systems: an organizational perspective", Addison-Wesley Publishing [2] Karacapidilis, N.I and Pappis, C " A framework for group decision support systems: Combining AI tools and OR techniques", European Journal of Operational Research, Vol. 103, pp , [3] Abdelkader ADLA "A Cooperative Intelligent Decision Support System for Contingency Management", Journal of Computer Science Vol.2, No.10, pp , [4] Roger L. Hayen, "Investigating decision support system frameworks", Issues in Information Systems Journal, Vol. 2, No.2, [5] Backus, M.,"E-Governance and Developing Countries, Introduction and examples", Research Report, No.3, 2001 [6] Sharma, S. K. and Gupta, J. N. D., "Building Blocks of an E- Government A Framework", Journal of Electronic Commerce in Organization, Vol.1, No.4, pp 34-48, [7] P.Salhofer, and D.Ferbas, "A pragmatic Approach to the Introduction of E-Government", Proc.8, International Government Research Conference, [8] Zhiyuan Fang, "E-government in Digital Era: Concept, Practice, and Development", International Journal of the Computer, the Internet and Management," Vol.10, No.2, pp 1-22, [9] Shivakumar Kolachalam, "An Overview of E- government",international Symposium on learning Management and Technology Development in the Information and Internet Age, online available at [10] Shailendra C. Jain Palvia, Sushil S.Sharma ", E.Government and E-Governance: Definitions/Domain Framework and Status around the World", 5 th International Conference on E-governance (ICEG), the patent No. 2003E DE HOL / NUR, Magnetic Resonance, SIEMENS Company, Erlangen, Germany, Furthermore, he is associate editor for journal of computer science and network security (IJCSNS) and journal of convergence in information technology (JCIT). Furthermore, he is a referee for IEEE Transactions on Signal Processing, Journal of Applied Soft Computing, the International Journal of Machine Graphics & Vision, the International Journal of Computer Science and Network Security, Enformatika Journals, WSEAS Journals and many different international conferences organized by IEEE. In addition, he has been awarded the Japanese Computer & Communication prize in April Moreover, he has been selected in who is who in Asia 2007 and BIC 100 educators in Africa Gamal H. El-Adl is lecturer assistant at Information Systems department in the faculty of Computer and Information Sciences in Mansoura University. Has been completed his B.S. degree in Information Systems from Mansoura University in 2004 and his master degree in 2008 under title " Watermarking of Relational Databases" and Main research points currently are databases, e- government, and software development. A.M. Riad is the head of information systems department at Faculty of Computers and Information Systems in Mansoura University. Has been completed his B.S. degree in electrical engineering from Mansoura University in 1982, his master degree in 1988 and doctoral degree in 1992 in electrical engineering. Main research points currently are intelligent information systems and e- government. Hazem M. El-Bakry (Mansoura, EGYPT ) received B.Sc. degree in Electronics Engineering, and M.Sc. in Electrical Communication Engineering from the Faculty of Engineering, Mansoura University Egypt, in 1992 and 1995 respectively. Dr. El-Bakry received Ph. D degree from University of Aizu - Japan in Currently, he is assistant professor at the Faculty of Computer Science and Information Systems Mansoura University Egypt. His research interests include neural networks, pattern recognition, image processing, biometrics, cooperative intelligent systems and electronic circuits. In these areas, he has published more than 55 papers in major international journals and 120 papers in refereed international conferences. Dr. El-Bakry has

58 38 Real-time Error Measurement System for MVB Protocol Su Goog Shon 1 and Soo Mi Yang 2 1 Department of Information and Telecommunication Engineering, The University of Suwon, Hwasung-city, Gyunggi-do, , Korea 2 Department of Internet Information Engineering, The University of Suwon, Hwasung-city, Gyunggi-do , Korea Abstract Recently, there are lots of control equipments in a train such as traction control, air conditioners and even internet access. For this reason, vehicle network must allow for the big amount of transmission data and must ensure the high reliability. After investigating about characteristics of multifunction vehicle bus, an error detection and analysis system is proposed. The proposed error analysis system can be used to verify high reliability of data transmission over multifunction vehicle bus. We explain how to implement the embedded error analysis system based on an ARM processor. The proposed system can detect all kinds of errors that IEC standard refers to. Keywords: Reliability, Measurement, MVB, TCN, IEC, Linux, TCP/IP 1. Introduction The recent challenge for railway industry is to make automatic electronic coupling of the vehicles through a data communication network. Over the last several years, a modern, versatile communication system on board trains, both to interconnect equipment located inside a railway vehicle and to allow communication between different vehicles, has been studied. At the train level, the data communication system should configure itself when vehicles are coupled and interconnected on the track. At the vehicle level, manufacturers should assemble pre-tested units, such as doors or air conditioners manufactured by subcontractors. Manufacturers could reduce development costs by utilizing one standard. Train operators could reduce spare parts and simplify maintenance and part replacement. Automatic electronic coupling of the vehicles with standardization could improve train safety and maintenance. An international standardization of data communication has been studied at both the train and vehicle levels. Especially in Europe, it is important to ensure cross-border traffic by standardizing track profiles, pneumatic hoses, traction voltages, operating procedures, and so on. Trains need a standard form of data communication for train control, diagnostics, and passenger information. Such a data communication network was specified by International Electro-technical Committee (IEC) as the Train Communication Network (TCN). The IEC groups worldwide deputies from over 20 countries worked several years within the IEC s Working Group 22 on the definition of the Train Communication Network [1][2]. The TCN was adopted as the international standard IEC in 1999 [3]. Train Communication Network is a real-time data network proposed for use on trains, consists of two different networks with somewhat different protocols [4], and is called the distributed control system. The TCN architecture also addresses all relevant configurations found in rail vehicles. It comprises the train bus connecting the vehicles and the vehicle bus connecting the equipment aboard a vehicle. The train bus architecture is divided into a Wire Train Bus (WTB) interconnecting all vehicles, and in each vehicle a MVB (Multifunction Vehicle Bus), as according to the TCN standard. The general architecture of a train communication network is shown in Figure 1 from [4].

59 39 reduction in the amount of cabling and increased reliability with respect to conventional wiring. Fig. 1 Train Communication System Error detection is a crucial part of the train safety and maintenance. The train network must satisfy minimum levels of requirements for message frame transmission integrity. The integrity simply means that a sufficient number of uncorrupted frames are delivered to perform target functions. The TCN is relatively known as having excellent error detection properties [5][6], even though there is no error detection scheme that can detect all possible errors. It is important that train operators or maintenance staff should note whether any error has occurred. Currently, error analyzers for MVB are based on signal or frame levels. This paper is interested in frame-based error analyzers that basically can capture and analyze frames over the MVB bus and can tell whether errors have occurred. The staff of a maintenance department is of particular interest about how many errors do occur over the MVB. Even though there are some commercial error analyzers available from Siemens, Duagon, etc, they still have some problems, such as missing error status information due to slow communication speed of RS- 232C, or inconvenience due to very short communication length of PCI or PCMCIA. In this paper, we propose a fast and user-friendly error analyzer that can detect transmission errors of the data communication system on a train. So, the staff can easily obtain the results of error analysis via Ethernet, which is fast and lengthy relatively. Also, characteristics of frames are reviewed on the MVB in a train and the error coding scheme is discussed. The error measurement system corresponding with IEC standard is designed and implemented based on an ARM processor and Linux operating system. It can capture all kinds of corrupted frames on the MVB and monitor the line status. We explain the internal architecture of our approach. Finally, we draw conclusions from the actual implementation. 2. MVB Controller The MVB protocol is used for connecting equipment within a single vehicle (e.g., a rail car) or within different vehicles in closed train sets. Each vehicle with MVB has its own vehicle bus connecting on board equipment, such as sensors and actuators for doors, brakes and air conditioning system. The MVB enables considerable The MVB is a standard communication medium to transport and exchange data among attached nodes. These nodes, which are physically connected to the bus, may vary in function, size, and performance at a physical layer level. Figure 2 shows a diagram of MVBC (Multifunction Vehicle Bus Controller) made by ADtranz [7], which is possible to be an MVB class 2 or a higher device that can have a processor and may exchange data. For the MVB data transport, it requires the MVBC that handles the bus traffic and arbitration without participation of external application CPU. The MVBC and CPU can exchange data through a common memory called the traffic store. The MCU in the MVBC of Figure 2 coordinates all data transfers between the MVB and the traffic store. Fig. 2 MVBC ASIC by ADtranz All information and data pertaining to the MVBC are found in the traffic store (memory). The traffic memory stores two types of data: process variables and messages that TCN buses transport. Process variables reflect the train s state, such as speed, motor current, and operator s commands. Message data carry infrequent but possibly lengthy information, for instance, diagnostics or passenger information. Message length varies between a few bytes to several kilobytes. The traffic memory is a shared memory to interface the MVB with the application CPU. This store is visible to both application CPU and MVBC. Devices with MVBC are operated on maser or slave mode. A port is used to provide a means to control access, data storage and interrupts such that data is read consistently when it is accessed simultaneously by the bus and by the application CPU. This requires at least two memory pages: one is accessed for the bus access, the other for application CPU. Each device owns a certain number of ports configured. Two kinds of ports exist: physical and logical ports. The traffic memory is divided into 4 partition areas such as logical address space, device address, service area, etc. In this paper, memory configuration mode 3 is used, where the port index table maps the logical port addresses to the device address of the traffic memory which are

60 40 specified in the master frames. The Port Control and Status Register (PCS) contain all the relevant information pertaining to one port. This information is used by MVBC to determine how it should handle the related port. The PCS includes the following information: port related information (function code, port description, event, etc.), data consistency check, telegram report, transfer acknowledge bits, and check sequences. It is important for the MVB error analyzer to enable the loading data from the PCS. In addition, the two interrupt request signals must be connected directly or via an interrupt controller to the CPU. 3. MVB Characteristics Transmission errors can occur due to corruption from network transmission noise. Detecting every possible corrupted frame is inherently impossible because any detection technique cannot find out all sets of bit errors. MVB adopts the Manchester encoding technology according to the IEC The fixed frequency is 1.5Mbps, and there is an 8-bit cyclic redundancy code (CRC) following each 64-bit frame data. The frame of MVB can be divided into two parts: master and slave frame, which consist of START, DATA, CRC and END. The general formats of frames on the MVB are shown in Figure 3 and discussed first. Frames start with a start delimiter preamble of 9 bits. Frame data includes from one to four data payload sections, with each payload being 16, 32, or 64 bits in size. Frames with more than 64 bits of data are broken into multiple 64-bit data payloads as shown. Hamming distance of 8 is obtained from the combined operation of CRC and Manchester coding. The MVB uses well designed error coding schemes. The MVB has the ability to detect two types of errors caused by noise during transmission: invalid delimiter encoding and check sequence values. The mechanism to detect errors in MVB is based on observing Manchester bit encoding errors and detecting mismatches of CRC between sent and received. The MVB uses two kinds of telegrams: the master and the slave frame. The master frame is a 16-bit word which consists of an F-Code (4bits) and an address (12 bits). All buses pertaining to the TCN provide two basic medium accesses: periodic (for data like process variables) and sporadic (for on-demand data traffic, such as messages). Periodic and sporadic data traffic share the same bus, but devices treat each separately. Figure 4 shows the TCN basic period. One device acting as master controls periodic and sporadic data transmission, which guarantees deterministic medium access. To accomplish this, the master alternates periodic and sporadic phases. Traffic is divided into basic periods of fixed duration either 1 or 2 ms on the MVB. At the start of a period, the master polls the process variables in sequence during a certain time period the periodic phase. To reduce traffic, urgent data are transmitted every period and less urgent variables are transmitted with an individual period every second, forth, eight, and so on basic period, with the longest period being 1,024 ms. Fig. 3 MVB message formats. Each data payload section is protected by an 8-bit Check Sequence (CS). The end of each frame is denoted by a 2- bit End Delimiter sequence. Frame length is inferred from the detection of an End Delimiter. The MVB uses a Check Sequence protecting every data payload segment of 16, 32, or 64 bits. The Check Sequence use the CRC polynomial for MVB, which is given by (1) G ( x) x x x x 1. (1) This polynomial guarantees a Hamming distance of 4. The CS encoding used by the MVB is known to detect successfully 1-bit and 2-bit errors [7]. The overall Fig. 4 MVB telegram Decoder accepts Manchester signals, strips delimiter and stores received data in receive buffer. Error information is transferred to the telegram analysis unit, where the information is processed. Some are passed to the Main Control Unit (MCU) for further actions. Telegram analysis unit of the MVBC plays key role to detect sorts of errors. Telegram analysis unit which consists of four registers: Frame Counter, Error Counter, Master Frame Register, and Master Frame Reg. Duplicate Exception handles the following tasks: reporting incoming slave frames, timeout mechanisms, telegram error handling, master frame registers, and telegram error recordkeeping. For example, the interrupt Slave Frame Checked is asserted by the announcement of the decoder that a valid or erroneous Slave Frame has arrived. The numbers of frames sent and received are recorded in the Frame Counter Register (FC) for obtaining quantitative

61 41 results regarding the quality of the bus. The number of erroneous frames sent or received is recorded in the Error Counter Register (EC). When 65,535 frames have been received, the interrupt Frames Evaluated Interrupt will be asserted [7]. Table 1 shows types of error that telegram analysis unit can detect and its register name used as in IEC The column of implementation in Table 1 means that each type of errors can be detected by using the proposed error detector. Table 1. Error types and registers Types of error Types of Implementation register Total Frame counter FC 0 Master Frame check MFC 0 Slave Frame check SFC 0 Erroneous Master Frame EMF 0 Erroneous Slave Frame ESF 0 Duplicate master frame DMF 0 Duplicate slave frame DSF 0 Bus timeout interrupt BTI 0 Reply timeout interrupt RTI 0 Frame evaluated interrupt FEV 0 Data transfer interrupt MCU 0 Tx queue exception TQE 0 Rx queue exception RQE 0 Transmit queue complete TQC 0 Receive queue complete RQC 0 Externel Interrupt 0-3 EI 0 All master frame transmitted AMFX 0 node. The MVB error detector is connected between the master and slave MVB node. Both the MVB master and the slave send and receive frames over the line. Then, frames captured from the MVB error detector pass through a decoder, a receive buffer, and a telegram analysis unit of the MVB. Whenever there is an error referred in Table 1, an interrupt event is asserted. Error analyzer on the Windows PC gets the results of error analysis system via TCP/IP communication channel and displays that on the screen. Fig. 5 MVB error analysis system Figure 6 shows the embedded MVB error detector implemented. The error detector mainly consists of MVB controller and CPU. The MVB controller is physically connected to MVB bus and is operated as one of MVB slave nodes. This node can only get command and data as a sink on the bus. The data, including all the process variables and messages, are stored into the traffic memory of 256KB interfaced to the MVB controller and are shared between the MVB controller and CPU (ARM). The ARM can access the traffic memory and its MVB data, and can transfer to the MVB error analyzer program on the PC via TCP/IP interface. One of the jobs that the ARM offers is to convert frame data of the MVBC into Ethernet format. 4. Implementation of MVB Error Measurement System Recently, there have been researches to develop embedded network equipment [8]. For the error analysis for the MVB frame transmissions, a MVB error analysis system is designed. The analysis system consists of both a MVB error detector and an error analyzer as shown in Figure 5. The MVB error detector can capture the MVB frames on the line and can tell what kinds of error occur or not. The MVB error analyzer can tell the results of error analysis for frames captured. MVB nodes are connected to the bus in series. Figure 5 shows an example system to gather MVB data for this paper. There is a maser MVB node and a slave MVB

62 42 M V B Fra me M V B Fr a ES M V B Fr a m RJ-45 Ethernet The MVB board driver is to process interrupt routine. It can initialize the traffic memory, store PCS information to the portlist, and register the PCS on the MVBC. Also, it can deal with the data communication between the traffic memory and DPRAM. Ethernet driver routine can offer the socket communication to the PC-based MVB error analysis program, where the socket program is operated on the non-blocking mode, and may connect to the PC program whenever it requires to connect. MVBC ARM Application program (Interrupt Traffic Memory access socket) Traffic memory Traffic data Linux (kernel 2.6.8) Fig. 6 Embedded MVB error detector Figure 7 describes the details for the ARM board. For this research, the Samsung 2410 of ARM 920T is used. The ARM 2410 is directly connected to MVB controller through ARM bus and gathers and process the MVB data. The MVBC handles the ARM access made to the traffic memory or internal registers in the MVBC. MVBC driver 5. Results Ethernet driver Fig. 8 Software platform for MVB error detector The proposed MVB error detector is implemented as shown in Figure 9. The MVB controller board is shown on the top and the ARM board is placed underneath it. The MVB error detector is connected between a master and a slave MVB node as shown in Figure 5 and can capture the MVB frames that move from the master to the slave and from the slave to the master. Fig. 7 Details of MVB ARM board Figure 8 shows the software platform for embedded MVB error detector. The system based on Linux kernel consists of MVBC driver, Ethernet driver, TCP/IP protocol stack, and application program that fetch, store, and process the traffic memory data. It uses the programming languages ANSI C for programming their devices. For this embedded system, a boot loader (uboot) is used to initialize the system and load the Linux kernel and so on. All the Linux kernel, Linux file system, drivers, and application programs are stored on NAND flash memory, and loaded into the SDRAM after booting. After booting, RS-232C or Ethernet interface is used for debugging and loading the application programs. Fig. 9 Implemented MVB error detector After booting the detector, the initialization process is shown in Figure 10. The MVB error detector starts to listen to both the bus and the Ethernet port. The initialization process is also shown by using the PCS registers information.

63 43 maintenance staff does not want to spend time installing or configuring anything in order to monitor their systems. They just want to perform maintenance. Acknowledgments This work was supported by the GRRC SUWON 2010-B5 program of Gyeonggi province. The advice received from K. H. Shon of Intercon system ltd is highly appreciated. Fig. 10 View of MVB error detector initialization Figure 11 shows the MVB error analysis program working on Windows with approximately 140KB binary application code. It is a simply GUI based socket program as shown in Figure 11. It can decide whether the frames are master or slave frames from the information for the F_Code. The error analysis program can tell the error analysis results, such as total frame, count, frame checked, reply timeout interrupt, etc, as listed in Table Conclusion Fig. 11 MVB error analysis program After investigating about MVB characteristics, an error detection and analysis system is designed and implemented. The error analysis system can show all the error information referred by IEC about PCS internal register value, error status, frame status, and link status after capturing MVB frames. This implementation contributes to the management and maintenance of the MVB node. It can offer to the maintenance people a user-friendly way to monitor all the equipment on a vehicle. Before this research, the error analysis was based on the RS-232C serial interface, which limits the speed of analysis. With the Ethernet interface, the error analysis can capture all the frames on the line now. For future work, we plan to develop MVB protocol analysis system with the automatic configuration because References [1] G.Fadin and F.Cabaliere, "IEC TC9 WG22 train communication network dependability and safety concepts, "World Congress on Railway Research 97, [2] H.Kirrmann and P.A. zuber, "IEC/IEEE train communication network." [3] IEC Standard Train Communication Network: Part (1) General Architecture (2) Real-time Protocol (3) Multifunction Vehicle Bus (4) Wire Train Bus (5) Train Network Management (6) Train Communication Conformance Testing, [4] UIC 556 Standard, Information Transportation on the Train Bus, [5] H. Kirrmann and P. A. Zuber. The IEC/IEEE Train Communication Network, IEEE Micro, 21(2):81 92, March/April [6] Philip Koopman, Analysis of the Train Communication Network Protocol Error Detection Capabilities, Technical Report, Carnegie Mellon University, Pttsburgh, PA, USA, Feb [7] ABB Daimler-Bensz Transportation (Switzerland) Ltd, "Multifunction vehicle Bus Controller," Adteanz, [8] Thomas Nolte, Share-Driven Scheduling of Embedded Networks, Malardalen University Press Dissertations No. 26, May, Su Goog Shon (corresponding author) received his B.S. (198 2) degree in Electrical Engineering from Seoul National Uni versity, his M.S. (1984) degree in Electrical Engineering fro m Seoul National University, and his Ph.D. (1996) degree i n Electrical and Computer Engineering from the University of Texas at Austin. He is an assistant professor in the Depa rtment of Information and Telecommunication at the Univer sity of Suwon in Korea. His research interests include comp uter and embedded system, network protocol, network sim ulation, and network programming. Soomi Yang received the B.S., M.S. and Ph.D. degrees in computer engineering from Seoul National University of Seoul, Korea, in 1985, 1987 and 1997 respectively. From 1988 to 2000, she was a researcher at Korea Telecom Research Center where she worked on telecommunication network, internet and information security. From 2000 to 2001, she was a visiting scholar at UCLA, USA. From 2002 to 2004, she was a faculty of the Suwon Science College. Since 2004, she has been on the Faculty of the University of Suwon, Korea, where she is a professor of computer

64 44 sciences. Her research interests in information security include access control, network security, and secure system software.

65 44 General Database Infrastructure for Image Retrieval Carlos Alvez 1, Aldo Vecchietti 2 1 Facultad de Ciencias de la Administración, Universidad Nacional de Entre Ríos Concordia, 3200, Argentina 2 INGAR UTN, Facultad Regional Santa Fe Santa Fe, S3002GJC, Argentina Abstract In this article, a general database infrastructure implemented in an Object-Relational Database Management System for image retrieval is proposed and describe. Semantic and content based image queries can be performed with this application. The infrastructure is structured into three levels: content-based, semantic data and an interface integrating them. It is complemented with a set of database User Defined Types (UDT) composed of attributes and operations. Set operations: union, intersection and difference are implemented for recovering images based on its attributes. In order to prove the capabilities of this approach a case study about vehicles is implemented. The results obtained by performing about 240 queries with a database image show an important improvement in image similarity search. The architecture can be easily adapted for specific field applications. Keywords: content semantic image retrieval - ORDBMS. 1. Introduction This paper is an extension and improved version of a previous one [1] where we introduced the software architecture to address image retrieval under an Object- Relational Database Management System (ORDBMS) [2]. Nowadays, the trend in image retrieval is the integration of both low-level and semantic data. Object-relational database management system (ORDBMS) and its standards (ISO/IEC , 2003; ISO/IEC , 2003) emerged at the end of the 90 s to incorporate the object technology into the relational databases, allowing the treatment of more complex data and relationships than its predecessors. One of these complex applications is Content-Based Image Retrieval (CBIR) which has received a great interest in the past decade. The interest has been driven by the need to efficiently manage and search large volumes of multimedia information, mostly due to the exponential growth of the World-Wide-Web (WWW). CBIR is performed based on abstract descriptions of the images that are extracted during the image analysis phase. Most of the techniques proposed for CBIR are limited by the semantic gap separating the low level information from its metadata annotations. Semantic gap for image retrieval is the difference between low level data extracted and the interpretation the user has for the same picture [3]. The use of semantic models and techniques generated for the World Wide Web has exponentially grown in the last years for many applications beyond the use for Internet. In order to follow this trend, databases have included this technology to expand the nature of the applications to be performed thorough the DBMS. In this work, Oracle Semantic Technologies included in its 11g ORDBMS version has been employed. Several approaches can be found in the literature about this topic. RETIN is a Search Engine developed by Gony et al. [4] with the goal of fulfilling the semantic gap. It uses an interactive process that allows the user to refine the query as much as necessary. The interaction with the user consists of binary levels to indicate if a document belongs to a category or not. SemRetriev proposed by Popescu et al. [5] is a prototype system which uses an ontology to structure an image repository in combination with CBIR techniques. The repository includes pictures gathered from Internet. Two methods are employed for image recovery based on keywords and visual similarities where the proposed ontology is used for both cases. Atnafu et al. [6] proposed several similarity-based operators for recovering images stored on relational databases tables. This is one of the first efforts to integrate the content-based and semantic for pictures in a DBMS. Tests were performed extending the prototype called EMIMS. The authors use a database as image repository and added several algorithms to EMIMS. Research in image retrieval has been handled separately in database management systems and computer vision. In general systems related to image retrieval assume some sort of system containing complex modules for processing images stored in a database.

66 45 In this version, several issues regarding the combined lowlevel and semantic features image retrieval are added. A deeper explanation about the architecture and its implementation is provided such that it can be used as a reference guide for this type of application. The article is outlined as follows: first the proposed architecture is presented and the levels composing it are explained with more detail. Then a vehicle case study is introduced where the reference ontology used for vehicle classification and MPEG-7 low-level descriptors selected for CBIR are described. After that, the considerations made to include semantic data to images are explained. Finally experiments performed, results obtained and conclusions are presented. 2. Architecture for image retrieval The architecture for semantic and content based image retrieval in an ORDBMS is shown Fig. 1. Three main layers compose this model: The low level tier is responsible of loading, extracting and management of images visual descriptors (color, texture, etc.). Combinations of user define type (UDT) and java functions and SQL queries are employed in this layer. The semantic level is composed of database interfaces to introduce semantic ontology concepts and to recovery images related to these concepts in response to SQL semantic queries. Between the previous layers exists the connection level which is the interface that links both levels. A special infrastructure of UDTs, functions and operators are provided in this section to relate the images low-level descriptors and its semantic concepts. Fig. 2 shows the UML class diagram representing the collection of database UDTs to support images, matrices, arrays of low level descriptors and semantic information. Classes of Fig. 2 are transformed (via mapping functions) to UDT when defining the ORDBMS schema based on SQL:2003 standard [2]. On top of the hierarchy there is an interface called SetOperations where the signature of set operators are defined. Those operators can be used to query low-level or semantic features separately or combined. Image is a class defined to store the image and its properties which inherits the operations declared in SetOperations. Image has a composition relationship with the generic LowLevelDescriptor abstract class from which special classes to implement specific low-level descriptors can be derived. Image attributes are: - a BLOB (Binary Large Object) to store pictures, - low-level descriptors UDTs type as a result of mapping the composition relationship with LowLevelDescriptor, - some other attributes to keep images characteristics like size, width, height, etc. Methods importimage and exportimage are responsible to import and export images from files respectively. The method similarscore implements a particular descriptor which can be used to obtain a distance between a reference image and another one stored in the database; similar method in Image class calls similarscore to compare the descriptor values of a particular image against the remaining in the database. Low-level UDTs are implemented according to the application domain. For example, for general images can be used descriptors proposed by the MPEG-7 standard, for more specific applications like biometry and medicine explicit domain descriptors must be employed. Set operators inherited from the interface can be redefined in Image UDT for special purposes. These UDTs and its operators facilitate CBIR queries by using SQL sentences which are familiar in the IT community. Fig. 1 Architecture proposed for image retrieval. SemanticData class is included to support domain ontology. It inherits methods from interface SetOperations. SemResultSet is the method to return a set of images with same semantic values. Semantic data are not loaded automatically and several alternatives are possible to work with them: metadata generated by the user, structured concepts extracted from Internet, ontology driven concepts, etc. Since semantic model must be structured in order to query them by its content, it is convenient to count with a domain ontology like the one proposed by MPEG-7 group [7].

67 46 ORDBMS and to get benefit of it s Semantic Technologies Tools [10]. 3.1 Low Level Implementation The implementation on this tier implies the creation of the UDTs with its attributes and methods. The low-level metadata for image recovery are represented as a collection of descriptors, precisely the LowLevelDescriptor UDT is created (Fig.3 Part A) where extractd and similarscore methods are declared. The implementation of these operators for a specific descriptor is defined in the subclasses (see section 4.1). Fig. 2 UML class diagram model of the object-relational database defined for the architecture. The composition relationship shown in Fig. 2 between image and LowLevelDescriptor is made through a collection, an array for this case because the collection multiplicity can be estimated. In Fig. 3 Part B, it is shown the descriptors_t UDT creation as an array of LowLevelDescriptor and in Fig. 3 Part C, in the definition of Image_t UDT the descriptors attribute is defined as of descriptors_t type. In this work, semantic is introduced for every image in the database. Concepts taken from the domain ontology or literals x are associated with a specific UDT Image instance y by means of the following triple: <subject property object> Links can be performed by asserted triples (<y rfd:type x>) or can be inferred by means of some other properties like subclass of, range, subproperty of, etc. [8][9]. The previous levels can be queried individually; in cases that low-level and semantic data are needed together the interface setoperations which is inherited by both is used to integrate and get results of combined queries. SetOperations lets the definition of combined union, intersection and difference set operations, which calls similarscore and semresultset functions of LowLevelDescriptor and SemanticData UDT respectively. By these methods it is also possible to get results from two low-level queries or two semantic queries. Both similarscore and semresultset functions returns a set of Image references (set<ref>). Since both methods return a set<ref> they can be easily employed in SQL queries. 3. Model Implementation The model implementation was made using Oracle 11g database. It was chosen because is one the most advanced Fig. 3 PL/SQL code. Part A: abstract UDT LowLevelDescriptor definition. Part B: UDT descriptors_t definition as array of LowLevelDescriptor. Part C: UDT Image_t attributes. Two alternatives are possible to define the descriptors amount and type to be used in a particular application. The first one is by defining the descriptors array inside Image_t constructor; and the second one is by defining an additional table to store the descriptors. The first

68 47 alternative was chosen for this case (Fig. 4 Part A and B). In Fig. 4 Part A can be seen the signature of the constructor method Image_t where filen, is a parameter which carries out the file name containing the image and filep corresponds to the file path and dirob is an Oracle Directory object needed to load the image. Fig. 4 Part B shows a code portion of the constructor method Image_t. In this method the descriptors array is created, then its particular instances and finally the arrays are loaded with data. In section 4.1 is shown the implementation of some of those descriptors. In the constructor method is also loaded the picture into a BLOB attribute. (Fig. 4 Part C). The Image_t UDT, has also the similar method (Fig. 5 Part A), which is a function that allows the similarity query of one image against the other stored in the database. This method returns a set of references to Image_t objects which are stored in its corresponding Type Table which is created by the following SQL sentence: CREATE TABLE Image OF Image _t; In order to get the image references it is necessary to define SetRef UDT as a set of Reference type to Image_t. Because the number of references that can be returned by the similar method is not known, the best way to implement SetRef UDT is by using a multiset collection which corresponds to nested tables in Oracle 11g (Fig. 5 Part B). similar has two parameters: the descriptor to be used (which is an index inside the descriptors array) and its threshold. For the image comparison all images in the database are invoked by calling the descriptor s similarscore method for each one (Fig. 5 Part C). Fig. 5 PL/SQL code. Part A: Function similar Signature. Part B: UDT SetRef definition. Part C: Function similar Image_t in Image_t s body. Fig. 4 PL/SQL code. Part A: Constructor Image_t Signature. Part B: Constructor Image_t in Image_t s body - Descriptors Extraction. Part C: Constructor Image_t in Image_t s body - Load Image.

69 Semantic Implementation In order to store and query the semantic information SemanticData class (Fig. 2) must be responsible for the following tasks: Import, load and store the application domain ontology. Load new concepts related to the ontology. Establish relationships between images and domain ontology concepts. Allow semantic queries to search images in the database. For SemanticData UDT implementations we employed the facilities provided by Oracle Semantic Technologies (OST). From Fig. 6, it can be seen that SemanticData_t UTD has two attributes: idtriple which corresponds to a triple identifier and the triple attribute which is of SDO_RDF_TRIPLE_S type predefined in OST. Given SemanticData_t UDT its corresponding Type Table is created as follows: CREATE TABLE SemanticData OF SemanticData_t; and then a semantic model must be created by means of OST: EXECUTE SEM_APIS.CREATE_RDF_MODEL('nom_model', 'SemanticTable', 'triple'); The first parameter corresponds to the model name, the second to the table name where semantic data are stored and the third is the attribute name where the triple must be stored OST also provides a Java API to import the ontology which was used for this case. For loading the semantic concepts and relate them to images instances we employed a constructor of SemanticData UDT, as follows: Insert into SemanticData Values( SemanticData_t( idtr, SDO_RDF_TRIPLE_S( stridimg, property, object), SemanticData_t( idtr, SDO_RDF_TRIPLE_S( stridimg, Rdf:type, :Image) ); The first inserted tuple associates the image id with a property and object of the domain ontology. The second inserted tuple indicates that the inserted id is of image type. This is important because when recovering images responding to the following query pattern <?I p O> it must also satisfy the triple <?I p :Image>. Fig. 6 PL/SQL code. UDT SemanticData_t definition. The static function semresulset is defined to query images using semantic data. This function takes as input parameters a property and an object and returns a set of references to Image. To perform this task this function employs the SEM_MATCH method defined by OST. In the same way than similar of Image_t, this function returns an object of type SetRef which can be used as an input to union/intersection/diference functions (Fig. 2). 3.3 Connection Level Implementation Having the structure of Image_t and SemanticData_t UDTs, the combined queries integrating both aspects is now very simple to do employing the set operators mentioned before. Anyone of the set operators has the form of: Op(SetRef, SetRef): SetRef. Due to similar and semresultset methods return a SetRef type, then in the set operators any of the following combinations is valid: Op(similar, similar): SetRef Op(semResultSet, similar): SetRef Op(semResultSet, semresultset): SetRef The meaning of it is that it is possible to combine not only semantic data queries with low-level data but also lowlevel queries with different descriptors or semantic queries with different patterns. Since the operators have been defined as static functions they do not depend on a particular instance. They can be used independently of a particular application. Fig. 7 shows the implementation of set_union function, which is done in Image_t and also in SemanticData_t, the

70 49 other functions set_intersect and set_diference are done in a similar way. Fig. 7 PL/SQL code. Set_union function implementation. 4. Case Study For the case study vehicle images are used. To provide low-level descriptors for those images MPEG-7 standard descriptors: Dominant Color (DCD), Color Layout (CLD), Scalable Color (SCD) and Edge Histogram (EHD) are selected. In the experiment, vehicle images gathered from Flickr by means of Downloadr [11] are stored in the database. Tags and texts are used to obtain pictures from Flickr site. In section 4.1 there is an explanation about low-level descriptors selected and how data are obtained. Fig. 8 presents an excerpt of the vehicle ontology used in this example, images tags and texts have been introduced as concepts in the database like RDF triples. Section 4.2 describes with more details this issue. similarscore method was written in PL/SQL, while extract is a Java wrapper code linked to LIRe library defined by Caliph&Emir[12]. Extract are static methods performing the task of creating objects, obtain descriptors data and store the information in the corresponding arrays defined in the UDTs [13]. For example, extractionehd creates an instance of EdgeHistogramImplementation calls method getstringrepresentation and returns the 80 bins of this descriptor. The code is compiled and load in the database by means of Loadjava utility, by means of the following sentences: Loadjava u usuario/password JavaWrapper.class In order to use ExtractionEHD function in Oracle another wrapper written in PL/SQL is needed for each method in JavaWrapper.class as follows: CREATE OR REPLACE FUNCTION extractehd(imagen VARCHAR2) RETURN VARCHAR2 AS LANGUAGE JAVA NAME 'JavaWrapper.extractionEHD(java.lang.String) return java.lang.string'; Fig. 9 represents the implementation model of low-level data descriptors. In Fig. 10 it is shown a detailed implementation of the UDT EdgeHistogram_t as a subtype of LowLevelDescriptor. In Fig Part A, BinCounts is defined as an array of 80 positions which contains the attribute bincounts of EdgeHistogram_t UDT (Fig Part B). The description of the body of EdgeHistogram_t and its methods extractd and similarscore are shown in Fig. 10 Part C. The first one calls extractehd function defined in the PL/SQL Wrapper previously presented. Function similarscore calculates the difference in absolute values of the bincounts array between the object calling the function and the array of the input parameter. Fig. 8. Excerpt of vehicle reference ontology. 4.1 Low Level. Mpeg-7 descriptors Four UDTs subclases of LowLevelDescriptor are implemented in Oracle to load the following MPEG-7 standard descriptors: Dominant Color (DCD), Color Layout (CLD), Scalable Color (SCD) and Edge Histogram (EHD). Extract methods were defined to obtain descriptors data and similarscore for content based search. Fig. 9. Low-level description implementation model.

71 50 function. Besides the triples of Fig. 8, some other are defined to relate concepts with class instances, domains, etc. Every image loaded in database table Image has a link to the reference ontology by adding a triple in the form of <I p O>), which associates rows of Image table with the hierarchy shown in Fig. 8. Some other functions provided by Oracle Semantic Technologies can be employed after this step, for example SEM_APIS.CREATE_ENTAILMENT creates a rule index to perform RDFS, OWLPRIME, user rules inference, or SEM_MATCH to query semantic data directly. 4.3 Low-level and Semantic Integration Implementation Results returned by semresultset y similarscore are set of image references that can be combined and manipulated by defining and implementing functions of the interface SetOperations. These operations can be overloaded and/or overridden for special purposes giving several options and flexibility to develop an application. 5 Experiments and Results A database of medium size containing images gathered from Flickr has been generated to perform the analysis, where 30 of them are proposed like descriptors query patterns. For each one between 15 and 20 images visually similar are proponed as the ground truth set (gts) [14]. The recovery rate (RR) can be obtained as follows: Fig. 10 PL/SQL code. Part A: BionCounts_t type. Part B: UDT EdgeHistogram_t definition. Part C: EdgeHistogram_t body. NF(, q) RR( q) NG( q) (1) Database type table Image is defined to store objects of Image_t UDT. Each table row contains data items of lowlevel descriptor UDTs such that once an image is loaded the appropriated extractd method is invoked to initialize the objects. 4.2 Semantic. Vehicle model reference The reference ontology was generated having in mind tags used to download the images from Flickr. It categorizes different vehicle types. By means of Oracle Semantic Technologies a semantic data network was created and also a RDF Vehicle model. These tasks are completed by implementing SemanticData UDT and semresultset where NG(q) is the gts size for query q and NF(, q) is the amount of images of the gts obtained between the first NG(q) recovered. RR(q) takes values between 0 and 1, if the value is 0 no images have been found from the gts while 1 indicates that all images of the gts were recovered. The factor must be 1, tolerance is larger for higher and therefore it increases the likely of obtaining images from the gts. For the experiment we set = 1.6 y NG(q) = 5 meaning that NF(, q) is the amount of gts images recovered from the first 8. Figure 11 presents an example and the gts proposed.

72 51 The performance of selected descriptors for all queries (NQ=30 for our study) is evaluated by the average recovery rate (ARR) [14] given by: 1 NQ ARR RR( q) NQ q1 (2) shown in the first and second series of Fig. 12, respectively. From Fig. 12 can be concluded that SCD descriptor has the best behavior when recovering images by similarity. It can be also observed that there is an important improvement getting similar images when semantic inference is employed together low level descriptors data. In Figure 13 it can be seen the images obtained with queries involving SCD descriptor integrated with semantic data. Fig. 12. ARR values obtained with queries performed. Fig. 11. Examples of ground truth set. Different levels of the hierarchy tree of Fig. 8 are selected as query objectives in order to compare the efficiency and precision of the answers. Results obtained taking into account with and without the semantic models are shown in Fig. 12. For example for two doors sedan cars the comparison is made considering the whole set of car images against 2_doors cars set. The inference made by using specified domain and property rdfs:subclassof can contain any of these values: 2_door, convertible, coupé. Images qualified by 2_doors concept are visited and also those belonging to coupe and convertible. The experiment is performed as follows: One query for every image in the database using each one of the descriptors presented in section 3.1. One query for every image satisfying the condition of the tree node selected in a direct or inferred manner. In total 240 queries have been executed. In Fig. 12 can be found the average recovered rate obtained for each one of the descriptors selected for this experiment, results without and including the semantic are Fig. 13. Results obtained with SCD descriptor. The left most image of each row shows the query image, the other four on the right show the most relevant retrieved.

73 52 6 Conclusions Architecture for image retrieval in an Object-Relational Database Management System is proposed and implemented in Oracle 11g database by using its semantic technologies framework. The architecture is structured in three levels, one for the semantic content of the images, another one for the low-level content and the third tier is the connection between the previous. For each layer, UDTs containing attributes and operator functions for images are proposed and defined. These UDTs are the fundamental infrastructure in order to support images, its characteristics and perform queries based on its low-level contents and semantic information. A detailed explanation about the implementation is made through this article to provide and insight about the way to carry out this application type. The architecture functionality has been studied by means of a case study based on a vehicle ontology for semantic annotations and MPEG 7 low-level descriptors for lowlevel contents. Descriptors employed are: Dominant Color (DCD), Color Layout (CLD), Scalable Color (SCD) and Edge Histogram (EHD). Extraction and loading data about those descriptors is made via Java and PL/SQL wrappers linked to LIRe library. Semantic data is included by means of RDF triples using Oracle tools. Ground truth sets of images and the average recovery rate (ARR) property are defined to drive the experiments. Semantic data of images have been loaded in the database using reference ontology concepts and properties. A total of 240 queries have been executed. Results obtained from the scheme proposed demonstrate an important improvement in images similarity search. This example belongs to a broad image domain, having an unlimited and unpredictable variability of the image's content. On the other hand, a narrow image domain can be easily implemented because the architecture proposed is generic such that it can be adapted to specific field applications. Acknowledgments Authors thanks to Universidad Tecnologica Nacional (UTN) and Universidad Nacional Entre Ríos (UNER) for the financial support received. References [1] Carlos E. Alvez, Aldo R. Vecchietti. Combining Semantic and Content Based Image Retrieval in ORDBMS. Knowledge-Based and Intelligent Information and Engineering Systems Lecture Notes in Computer Science, 2010, Volume 6277/2010, Editors Rossitza Setchi. Ivan Jordanov, Robert J. Howlett, Lakhmi C. Jain. Springer- Verlag Berlin Heidelberg (2010). [2] Melton Jim, (ISO-ANSI Working Draft) Foundation (SQL/Foundation), ISO/IEC :2003 (E), United States of America (ANSI), [3] Neumamm D. and Gegenfurtner K. Image Retrieval and Perceptual Similarity ACM Transactions on Applied Perception, Vol. 3, No. 1, January 2006, Pages [4] Gony J., Cord M., Philipp-Foliguet S. and Philippe H. RETIN: a Smart Interactive Digital Media Retrieval System. ACM Sixth International Conference on Image and Video Retrieval CIVR July 9-11, 2007, Amsterdam, The Netherlands, pp , (2007). [5] Popescu A., Moellic P.A. and Millet C. SemRetriev an Ontology Driven Image Retrieval System. ACM Sixth International Conference on Image and Video Retrieval CIVR July 9-11, Amsterdam, The Netherlands, pp , (2007). [6] Atnafu S., Chbeir R., Coquil D. and Brunie L. Integrating Similarity-Based Queries in Image DBMSs ACM Symposium on Applied Computing, March 14-17, 2004, Nicosia, Cyprus, pp (2004). [7] J. Hunter, "Adding Multimedia to the Semantic Web - Building and Applying an MPEG-7 Ontology", chapter of "Multimedia Content and the Semantic Web: Methods, Standards and Tools", Wiley, [8] Dan Brickley and R.V. Guha. RDF Vocabulary Description Language 1.0: RDF Schema. W3C Recommendation 10 February 2004, [9] Allemang D. and Hendler J., Semantic Web for the working Ontologist. Effective Modeling in RDFS and OWL. Morgan Kaufman(2008). [10] Chuck Murray, Oracle Database Semantic Technologies Developer's Guide. 11g Release 1 (11.1) Part B September [11]Flickr Downloadr 2.0.4, [12] Lux Mathias, Savvas A. Chatzichristofis. Lire: Lucene Image Retrieval An Extensible Java CBIR Library. In proceedings of the 16th ACM International Conference on Multimedia, pp , Vancouver, Canada, [13] C. Alvez, A. Vecchietti, "A model for similarity image search based on object-relational database", IV Congresso da Academia Trinacional de Ciências, 7 a 9 de Outubro de Foz do Iguaçu - Paraná / Brasil (2009). [14] Manjunath B., Ohm J. R., Vasudevan V. and A. Yamada. Color and texture descriptors. IEEE Trans. on Circuits and Systems for Video Technology, vol. 11, no. 6, , June C. Alvez obtained his M.Sc degree in computer science in 2003, Departamento de Ciencias e Ingeniería de la Computación Universidad Nacional del Sur. Bahía Blanca, Argentina. Professor and researcher at the College of Administration Sciences, UNER. A Vecchietti. Obtained his PhD degree in 2000 in Chemical Engineering at Facultad de Ingeniería Química Universidad Nacional del Litoral, Argentina. He is a researcher of CONICET and Professor at Universidad Tencnológica Nacional, Facultad Regional Santa Fe, Argentina.

74 53 New approach using Bayesian Network to improve content based image classification systems Khlifia Jayech 1 and Mohamed Ali Mahjoub 2 1 SID Laboratory, National Engineering School of Sousse Sousse, Technology Park 4054 Sahloul, Tunisia 2 Preparatory Institute of Engineering of Monastir Monastir, Ibn Eljazzar 5019, Tunisia Abstract This paper proposes a new approach based on augmented naive Bayes for image classification. Initially, each image is cutting in a whole of blocks. For each block, we compute a vector of descriptors. Then, we propose to carry out a classification of the vectors of descriptors to build a vector of labels for each image. Finally, we propose three variants of Bayesian Networks such as Naïve Bayesian Network (NB), Tree Augmented Naïve Bayes (TAN) and Forest Augmented Naïve Bayes (FAN) to classify the image using the vector of labels. The results showed a marked improvement over the FAN, NB and TAN. Keywords: Bayesian Network, TAN, FAN, Image classification, Recognition, CBIR. Our paper presents a work which uses three variants of naïve Bayesian Networks to classify image of faces, using the structure of dependence found between objects. This paper is divided as follows: Section 2 presents the principal works related to CBIR and a state of art of 2D face recognition; Section 3 describes the developed approach based in Naïve Bayesian Network, we describe how the spatial features are extracted; and we introduce the method of building the Naïve Bayesian Network, Tree Augmented Naïve Bayes (TAN) and Forest Augmented Naïve Bayes (FAN), and inferring posterior probabilities out of the network; Section 4 presents some experiments; finally, Section 5 presents the discussion and conclusions. 1. Introduction In the few last years, the domain of indexation by the content called CBIR (Content Based Image Retrieval), which includes all that works developed to search and classify images through their internal content, has been a very active domain of research. Much of the related work on image recognition and classification for indexing, classifying and retrieval has focused on the definition of low-level descriptors and the generation of metrics in the descriptor space [1]. These descriptors are extremely useful in some generic image classification tasks or when classification is based on query by example. However, if the aim is to classify the image using the descriptors of the object content this image. There are two questions to be answered in order to solve difficulties that are hampering the progress of research in this direction. First, how shall we semantically link objects in images with high-level features? That means how to learn the dependence between objects that reflect better the data? Second, how shall we classify the image using the structure of dependence finding? 2. Related works 2.1 State of the art in 2D Face Recognition Face recognition is a huge research area and is the preferred mode of identity recognition by humans: it is natural, robust. Each year, the attempted solutions and algorithm grow in complexity and execution time. Zhao et al. in [28] divide the existing algorithms into three categories, depending on the information they use to perform the classification: appearance-based methods (also called holistic), feature-based and hybrid. Sharavanan et al. in [42] present taxonomy of face recognition methods used as shown in the figure 1.

75 Image based face recognition Holistic methods Feature methods Model methods Hybrid methods based based PCA LDA ICA Elastic bunch matching graph Dynamic link matching Active Model Appearance 3D Morphable Models Markov Random Field methods Independent Component Analysis Independent component analysis (ICA) is a generative model, expressing how observed signals are mixed from underlying independent sources. This approach transforms the observed vector into components which are statically maintained as independent. The basic restriction is that the independent components must be non-gaussian for Independent Component Analysis to be possible. A limitation of holistic matching is that it requires accurate normalization of the faces according to position, illumination and scale. Variations in these factors can affect the global features of the face, leading inaccurate final recognition. Moreover, global features are also sensitive to facial expressions and occlusions. Holistic methods such as neural networks [24] are more complex to implement, whereby an entire image segment can be reduced to a few key values for comparison with other stored key values with no exact measures or knowledge such as eye locations. 54 Fig. 1 Taxonomy of face recognition methods used [42]. In this section, we review the relevant work on 2D face recognition, and discuss the merits of different representations and recognition algorithms Holistic Approach Holistic algorithms use global features of the complete face. Examples of holistic systems include Principal Component Analysis (PCA) [31], Linear Discriminant Analysis (LDA) [33] as well as Independent Component Analysis (ICA) [32]. These projection techniques are used to represent face images as a lower-dimensional vector, and the classification itself is actually performed by comparing these vectors according to a metric in the subspace domain. Principal Component Analysis Principal Component Analysis (PCA) is an approach to analyze and dimensionally reduce the data, highlighting their similarities and differences. The task of facial recognition is discriminating input signals (image data) into several classes (persons). Linear Discriminate Analysis Linear Discriminate Analysis is effective to encode discriminatory information. The idea of LDA is grouping similar classes of data where as PCA works directly on data. It tries to find directions along which the classes are best separated by taking into consideration not only the dispersion within-classes but also the dispersion between classes Feature-based Approach Feature-based algorithms use local features [26] or regions [27] of the face for recognition. Such features may be geometric measurements, particular blocks of pixels, or local responses to a set of filters. Feature-based algorithms try to derive an individual s face s model based on local observations. Examples of such systems include the Elastic Bunch Graph Matching (EBMG) [38], recent systems using Local Binary Patterns (LBP) [34] [37], and also statistical generative models: Gaussian Mixture Models (GMM) [35] [36]. Elastic Bunch Graph Matching This technique use the structure information of a face, which shows the fact that the images of the same subject tend to deform, scale, translate, and rotate, in the plan of image. It uses the labeled graph, edges are labeled according to the distance information and nodes are labeled with wavelet coefficients in jets. This graph can be after used to produce image graph. The model graph can be deformed, scaled, translated, and rotated during the matching process. This can make the system resistant to large variation in the images. Dynamic Link Matching In dynamic link structure, the models are presented by layers of neurons, which are labeled by jets. Jets are wavelet elements describing grey level distribution. This approach encodes information using wavelet transformations. Dynamic link structure establishes a mapping between all-in-all linked layers, and thus reduces distortion. The method uses a winner-take-all strategy once a correct mapping is obtained choosing the right model. A typical dynamic link structure is given in the figure 2.

76 55 approaches using a Markov Random Field (MRF) model. According to the proposed method, the image is divided into smaller image patches, each having specific Ids. The model can be represented as in the figure 3. It includes two layers, observable nodes (squares in the graph, representing image patches) and hidden nodes (circles in the graph, representing the patch Ids). Edges in the graph depict relationships among the nodes. Fig. 2 Dynamic Link Architecture [42] Face recognition systems using local features empirically show a better performance as compared to holistic algorithms [40][36]. Moreover, they also have several other advantages: first, face images are not required to be precisely aligned. This is an important property, since it increases robustness against imprecisely located faces, which is a desirable behavior in real-world scenarios. Second, local features are also less sensitive to little variations in illumination, scale, expressions and occlusions. To go beyond the limitations of the Eigen faces method in holistic methods, Garcia et al. in [30] present a new scheme for face recognition using wavelet packet decomposition. Each face is described by a subset of band filtered images containing wavelet coefficients. These coefficients characterize the face texture, and a set of simple statistical measures permits forming compact and meaningful feature vectors. Then, an efficient and reliable probabilistic metric derived from Bhattacharrya distance is used in order to classify the face feature vectors into person classes. The Wavelet transformation allows the generation of such facial characteristics that are invariant to lighting conditions and overlapping. It organizes the image in subgroups that are localized according to orientation and frequency. In every subgroup, each coefficient is also localized in space Hybrid Approach Hybrid matching methods use a combination of global and local-features for recognition [33]. This approach is most effective and efficient in the presence of irrerelevant data. The key factors affecting performance depend on the selected features and the techniques used to combine them. Feature based and holistic methods are not devoid of drawbacks. Feature based method is sometimes badly affected by accuracy problem since accurate feature localization is its very significant step. On the other hand, holistic approach uses more complex algorithms demanding longer training time and storage requirements. Huang et al in [41] proposed a hybrid face recognition method that combines holistic and feature analysis-based Fig. 3 MRF model [41] 2.2 State of the art in Bayesian classification There are several works in CBIR systems using Bayesian Network. The majority of them involve various small problems, such as structure of Bayesian Network, feature extraction, classification and retrieval. Rodrigues in [3] presents a Bayesian network model which combines primitives information of color, edge-map and texture to improve the precision in content-based image retrieval systems. This approach allows retrieving images according to their features. However the spatial relations between objects are not treated. Zhang in [2] suggested modeling a global knowledge network by treating and entire image as a scenario. Authors suggested a process that is divided into two stages: the initial retrieval stage and Bayesian belief network inference stage. The initial retrieval stage is focused on finding the best multi-features space, and then a simple initial retrieval is done within this stage. The Bayesian inference stage models initial beliefs on probability distributions of concepts from the initial retrieval information and construct a Bayesian belief network. In this approach, the results are largely depending on the choice of representative blocks in the first stage. However, the dependence between objects in the approach presented in [2] is also not dealt with. Mahjoub and Jayech in [4] investigated three variants of Bayesian Networks such as Naïve Bayes, Tree Augmented Naïve Bayes (TAN) and Forest Augmented Naive Bayes (FAN) in image classification. Those models study the dependence between objects. However, the images used in this study represent the structure of a document containing texts blocks and graphs. The work presented in this paper tries to accomplish the difficult task of classifying the images of faces using the high level features. Our works study the dependence between the high level features to build an optimal structure that better represents the data saying if an object

77 56 is depends or not to another by using three variants of naïve Bayesian Networks. 3. The proposed Approach Data base of Pretreatment and Features i Descrip Learning structure Learning C 1 C 2 C i C N P(C1 P(C2 P P Query Image (R) Infer Class Maxi P ( Ci R) Classifying an image depends on classifying the objects within the image. The studied objects in an image are the following analysis represented by blocks. An example of one image being divided into elementary building blocks that contain different concepts is illustrated in figure 4: 3.1 Features Processing Color Feature Color may be one of the most straight-forward features used by humans for visual recognition and discrimination. To compute this feature we used histogram of color. The histogram can be considered as a Gaussian Mixture Model [6] (figure 5) defined as: K fx θ p f x α Where: p is the proportion of class i (p 0 and p 1 α μ,, μ and were respectively the center and the covariance of the kième normal component f. α. The estimation of these parameters can be doing with the EM proposed by Dempster and al in 1977 and the number of k normal component can be estimated as follows: BICK 2ln fxθ K v K ln n The value of k is chosen between k=1 and k (k is to choose in priori) that minimizes the Bayesian criterion BIC [7]. θ K and v K are respectively the maximum credibility estimator and the degree of freedom of the model Probability density Gaussienne estimated by EM X Fig. 5 Approximation of histogram by GMM Texture Features Fig. 4 Example of an image divided into some elementary building blocks For each block, we compute the histogram of color using GMM (Gaussian Mixture Model) and we compute the descriptor textural using the Grey Level Co-occurrence Matrix (GLCM). Then, we use Kmeans to cluster the object into k cluster. So each block will be labeled and integrated to one of the k cluster Textural features are linked to the contrast of processed images. The textural features can be used to differentiate between objects having the same color. Among these features, we used the Haralick features [5]. Haralick features computing is based on co-occurrence matrix. From these features, we use four attributes which are: energy, entropy, contrast, and homogeneity. The energy is defined as: E pi, j, The entropy is defined as ENT pi, jlog pi, j

78 57 The contrast is defined as CONT i j pi, j The homogeneity is defined as HOM 3.2 Clustering with K-means, 1 pi, j 1i j The attributes of this work is the clustering of vectors descriptors of the images to improve the classification by Bayesian Network by optimizing its construction. For each object extracted of the image, we compute the two features describing a given object: color and textural features. The feature attributes are calculated for each image of the data-base and then clustered into k cluster. We use the method of k-means to cluster the descriptor as shown in figure 6. Tree Augmented Naïve Bayes (TAN) It is obvious that the conditional independence assumption in naïve Bayes is in fact rarely true. Indeed, naive Bayes has been found to work poorly for regression problems (Frank et al., 2000), and produces poor probability estimation(bennett, 2000). One way to alleviate the conditional independence assumption is to extend the structure of naive Bayes to explicitly represent attribute dependencies by adding arcs between attributes. Tree augmented naive Bayes (TAN) is an extended treelike naive Bayes (Friedman et al., 1997), in which the class node directly points to all attribute nodes and an attribute node can have only one parent from another attribute node (in addition to the class node). Figure 8 shows an example of TAN. In TAN, each node has at most two parents (one is the class node). Class Fig. 6 The result of clustering with k-means into five clusters 3.3 Classifiers and learning In this study, we utilize Naïve Bayes, Tree Augmented Naïve Bayes (TAN) and Forest Augmented Naïve Bayes (FAN) classifiers. This sub-section describes the implementation of these methods Structure learning Naïve Bayes Network (NB) A variant of Bayesian Network is called Naïve Bayes. It is one of the most effective and efficient classification algorithms. Figure 7 shows graphically the structure of naïve Bayes, each attribute node has the class node as it parent, but does not have any parent from attribute node. As the values of can be easily calculated from training instances, naïve Bayes is easy to construct. Class Fig.7 Naïve Bayes (NB) Fig. 8 Tree Augmented Naïve Bayes (TAN) Construct-TAN procedure is described here as a blend of Friedman et al. s and Perez et al. s work. The procedure uses conditional mutual information, between two features given the class variable, which is described as follows: Px, y z I A,A C Px, y, zlog Px zpy z A A C A and A are sets of feature variables and C is the set of class variable, and x, y and z are the values of variables X, Y and Z respectively. This function measures the information provided by A and A for each other when the class variable is given. Algorithm : Construct-TAN Require: Naïve Bayes, training set Ensure: Structure of TAN 1: Calculate the conditional mutual information I A,A C between each pair of attributes, i j. 2: Build a complete undirected graph in which nodes are attributes A,,A. Annotate the weight of an edge connecting A to A by I A,A C. 3: Build a maximum weighted spanning tree. 4: Translate the resulting undirected tree to a directed one by choosing a root attribute and setting the direction of all

79 58 edges to be outward from it. 5: Construct a TAN model by adding a vertex labeled by C and adding an arc from C to eacha. Forest Augmented Naïve Bayes (FAN) The experiments show that classification accuracy in TAN is higher than in Naïve Bayes. Two factors contribute to this fact. First, some edges may be unnecessary to exist in a TAN. The number of the edges is fixed to n-1. Sometimes, it might outfit the data. Second, in step 4, the root of the tree is randomly chosen and the directions of all edges are set outward from it. However, the selection of the root attribute is important because it defines the structure of the resulting TAN, since a TAN is a directed graph. It is interesting that the directions of edges in a TAN do not significantly alter modify the classification accuracy. So, we correspondingly modify the TAN algorithm. First, we remove the edges that have conditional mutual information less than a threshold. To our understanding, these edges have a high risk to outfit the training data, and thus undermine the probability estimation. More precisely, we use the average conditional mutual information Iavg, defined in Eq.(1), as the threshold. Iavg, I A ;A C (1) Where n is the number of attributes. Second, we choose the attribute A with the maximum mutual information with class, defined by Eq.(2), as the root: A argmax A I A ;C (2) Where i=1,,n. It is natural to use this strategy, since intuitively the attribute which has the greatest influence on classification should be the root of the tree Since the structure of the resulting model is not a strict tree, we call our algorithm Forest augmented naive Bayes (FAN), described in detail as follows. Algorithm : Construct-FAN Require: TAN, training set Ensure: Structure of FAN 1: Calculate the conditional mutual information I A,A C between each pair of attributes, i j, and calculate the average conditional mutual information I,defined as follows: Iavg, I A ;A C nn 1 2: Build a complete undirected graph in which nodes are attributes A,,A. Annotate the weight of an edge connecting A to A by I A,A C. 3: Search a maximum weighted spanning tree. 4:Calculate the mutual information I A C,i 1,2,,,n between each attribute and the class, and find the attribute A that has the maximum mutual information with class, as follows: A argmax A I A ;C 5: Transform the resulting undirected tree into a directed one by setting A as the root and setting the direction of all edges to be outward from it. 6: Delete the directed edges with the weight of the conditional mutual information below the average conditional mutual informationiavg. 7: Construct a FAN model by adding a vertex labeled by C and adding a directed arc from C to eacha Parameters learning NB, TAN and FAN classifiers parameters were obtained by using the procedure as follows (Ben Amor 2006). In the implementation of NB, TAN and FAN, we used the Laplace estimation to avoid the zero-frequency problem. More precisely, we estimated the probabilities Pc,Pa c and Pa a,c using Laplace estimation as follows. Pc Nc 1 Nk Pa c Nc, a 1 Nc v Pa a,c Nc, a,a 1 Nc, a v Where - N: is the total number of training instances. - k: is the number of classes, - v : is the number of values of attribute A, - Nc: is the number of instances in class c, - Nc, a : is the number of instances in class c and with A a, - Nc, a : is the number of instances in class c and with A a, - Nc, a,a : is the number of instances in class c and with A a anda a Classification In this work the decisions are inferred using Bayesian Networks. Class of an example is decided by calculating posterior probabilities of classes using Bayes rule. This is described for both classifiers. NB classifier In NB classifier, class variable maximizing Eq.(3) is assigned to a given example. PC A PCPA C PC PA C (3)

80 59 TAN and FAN classifiers In TAN and FAN classifiers, the class probability P (C A) is estimated by the following equation defined as: PC A PCPA A,C Where A is the parent of A and PA A,C Nc, a,a si A existe Nc, a PA A,C Nc, a si A Nc n existe pas The classification criterion used is the most common maximum a posteriori (MAP) in Bayesian Classification problems. It is given by: da argmax Pclasse A argmax PA classexpclasse 4. Experiments and Results N argmax PA classexpclasse Now, we present the results of the contribution of our approach to classify images of some examples of classes from the database used 'Database of Faces'. The ORL face database consists of ten different images of each of 40 distinct subjects with 92*112 pixels. For some subjects, the images were taken at different times, varying the lighting, facial expressions (open /closed eyes, smiling / not smiling) and facial details (glasses / no glasses) as shown at figure 9. All the images were taken against a dark homogeneous background with the subjects in an upright, frontal position (with tolerance for some side movement). 4.1 Structure learning We have used Matlab, and more exactly Bayes Net Toolbox of Murphy (Murphy, on 2004) and Structure Learning Package described in (Leray and al. 2004) to learn structure. Indeed, by applying the algorithm of TAN and FAN with different number of cluster, we obtained the structures as follow: Fig. 10 Structure of TAN obtained with number of cluster k=5 Fig. 11 Structure of FAN obtained with number of cluster k=5 and threshold S1 Fig.12 Structure of TAN obtained with number of cluster k=10 Fig. 9 Example 'Database of Faces'

81 Results Fig. 13 Structure of FAN obtained with number of cluster k=10 and threshold S1 Fig.14 Structure of FAN obtained with number of cluster k=10 and threshold S2 4.2 Parameter learning We estimated the conditional and a priori probability Pc,Pa c and Pa a,c using Laplace estimation. We have used the Laplace estimation described in section (3.3.2) to avoid the zero-frequency problem; we obtained the results as follows: Table 1: a priori probability of attribute class Pclass P (class) class class class class class ,2 0,2 0,2 0,2 0,2 Table 2: a priori probability of Pa c with number of cluster k=5 P (a1 c) c1 c2 c3 c4 c5 1 0,80 0,84 0,28 0,84 0,76 2 0,04 0,04 0,04 0,04 0,04 3 0,04 0,04 0,04 0,04 0,04 4 0,08 0,04 0,60 0,04 0,12 5 0,04 0,04 0,04 0,04 0,04 For each experiment, we used the percentage of correct classification (PCC) to evaluate the classification accuracy defined as: number of images correctly classified PCC total number of images classified Other criterion can be used, such as: PXC1 classc1 and PXC2 classc2 represent the rate of good classification. P(X=C1 class=c2) and P(X=C2 class=c1) represent the rate of bad classification. The results of experiments are summarized in Table 3 and Table 4 with different number of clusters k=5 and k=10. Naive Bayes, TAN and FAN use the same training set and are exactly evaluated on the same test set. Table 3: Classification accuracy of NB, TAN and FAN with number of cluster k=5 cl cl cl cla mean clas as as as ss classific s 1 s s s 2 ation NB TAN FAN FAN (new thres hold) train ing set test set train ing set test set train ing set test set train ing set test set 0,80 0,9 0 0,40 0,8 0 0,15 0,9 0 0,10 0,7 0 0,40 0,6 5 0,30 0,7 0 0,75 0,9 5 0,30 0,8 0 0, 75 0, 80 0, 40 0, 50 0, 30 0, 40 0, 85 0, 80 0, 60 0, 50 0, 05 0, 10 0, 05 0, 10 0, 70 0, 70 0, 55 0, 60 0, 05 0, 10 0, 25 0, 10 0, 60 0, 60 0,72 0,62 0,31 0,30 0,33 0,32 0,77 0,64 As shown in experiments results in Table 3, the Naïve Bayes performed better than Tree Augmented Naïve Bayes (TAN) and Forest Augmented Naïve Bayes (FAN). However, when we increase the threshold in FAN we find out that the rate of mean classification is improved and became better than NB and TAN. Table 4: Classification accuracy of NB, TAN and FAN with number of cluster k=10 cla ss 1 cla ss 2 cla ss 3 cla ss 4 cla ss 5 mean classific ation RN train ing set test set 0,9 5 0,4 0 1,0 0 1,0 0 0,8 0 0,5 0 0,9 5 0,6 0 0,8 5 0,8 0 0,91 0,66

82 61 TAN FAN FAN (new thresh old) train ing set test set train ing set test set train ing set test set 0,6 0 0,1 0 0,7 5 0,3 0 0,9 5 0,3 0 0,7 5 0,9 0 0,9 0 0,9 0 1,0 0 0,9 0 0,4 5 0,4 0 0,2 5 0,4 0 0,8 5 0,7 0 0,3 5 0,3 0 0,5 5 0,6 0 0,9 0 0,8 0 0,4 5 0,4 0 0,4 0 0,3 0 0,9 0 0,9 0 0,52 0,42 0,57 0,50 0,92 0,72 As shown in experiments results in Table 4, the classification accuracy with number of cluster equal to 10 performed better than classification accuracy with number of cluster equal to 5. So, we were obliged to study the effect of variation of number of cluster in the classification accuracy Fig.15 Variation of classification accuracy saved by Naïve Bayes with different number of cluster (k=5, 8, 10, and 15) As experiments results shown in figure 15, we see that the mean of classification accuracy saved by Naïve Bayes is higher when number of cluster k=8 and then it decreased. So, we concluded that number of cluster is optimal when k=8 that means the number of cluster in this case reflect better the data. 5. Conclusion class 1 class 2 class 3 class 4 class 5 k=5 k=8 k=10 k=15 In this study we have presented a new approach for classifying image of faces by using Bayesian Network and K- means to cluster the vectors descriptors of the images to improve the classification using Bayesian Network by optimizing its construction. We have implemented and compared three classifiers: NB, TAN and FAN. The goal was to be able to apply algorithms that can produce useful information from a high dimensional data. In particular, the aim is to improve Naïve Bayes by removing some of the unwarranted independence relations among features and hence we extend Naïve Bayes structure shown in figure 7 by implementing the Tree Augmented Naïve Bayes. Unfortunately, our experiments show that TAN performs even worse than Naïve Bayes in classification. Responding to this problem, we have modified the traditional TAN learning algorithm by implementing a new learning algorithm, called Forest Augmented Naïve Bayes. We experimentally test our algorithm in data image of faces and compared it to NB and TAN. The experimental results show that FAN improves significantly NB classifiers performance in classification. In addition, the results show that the mean of classification accuracy is better when the number of cluster is optimal that means the number of cluster that can reflect better the data. 6. References [1] A. Mojsilovic, A computational model for color naming and describing color composition of images, IEEE Transactions on Image Progressing, 2005, Vol.14, No.5, pp [2] Q. Zhang and I. Ebroul, A bayesian network approach to multifeature based image retrieval, First International Conference on Semantic and Digital Media Technologies, GRECE, [3] P.S. Rodrigues and A.A. Araujo, A bayesian network model combining color, shape and texture information to improve content based image retrieval systems, 2004 [4] M.A. Mahjoub and K. Jayech, Indexation de structures de documents par réseaux bayésiens, COnférence en Recherche d'infomations et Applications, CORIA, 2010, Tunisia, pp [5] S. Aksoy and R. Haralik, Textural features for image database retrieval, IEEE Workshop on Content-Based Access of Image and Video Libraries, June, 1998, USA. [6] C. Biernacki and R.Mohr, Indexation et appariement d images par modèle de mélange gaussien des couleurs, Institut National de Recherche en Informatique et en Automatique, Rapport de recherche, No.3600, Janvier, [7] E.Lebarbier and T.M.Huard, Le critère BIC: fondements théoriques et interprétation, Institut National de Recherche en Informatique et en Automatique, Rapport de recherche, No.5315, Septembre, [8] A. Sharif and A. bakan, Tree augmented naïve baysian classifier with feature selection for FRMI data, [9] N.Ben Amor, S. Benferhat, Z. Elouedi, Réseaux bayésiens naïfs et arbres de décision dans les systèmes de détection d intrusions RSTI-TSI, Vol. 25, No.2, 2006, pp [10] D.M.Chikering, Learning bayesian network is NP-complete, In Learning from data: artificial intelligence and statistics, pp ,new York, [11] J.Cerquides and R.L.Mantaras, Tractable bayesian learning of tree augmented naïve bayes classifiers, January, 2003 [12] J.F.Cooper, Computational complexity of probabilistic inference using bayesian belief networks, Artificial Intelligence, Vol.42, pp , 1990.

83 62 [13] N.Loménie and N.Viencent, R.Mullot, Les relations spatiales : de la moélisation à la mise en œuvre, Revue des nouvelles technologies de l information, cépadues-éditions,2008 [14] N.Friedman and N.Goldszmidt, Building classifiers using bayesian networks, Proceedings of the American association for artificial intelligence conference, [15] O.Francois, De l identification de structure de réseaux bayésiens à la reconnaissance de formes à partir d informations complètes ou incomplètes, Thèse de doctorat, Institut National des Sciences Appliquées de Rouen, [16] O.Francois and P.Leray, Learning the tree augmented naïve bayes classifier from incomplete datasets, LITIS Lab., INSA de Rouen, [17] F.Hsu, Y.Lee, S.Lin, 2D C-Tree Spatial Representation for Iconic Image, Journal of Visual Languages and Computing, pp , [18] G.Huang, W.Zhang, L.Wenyin, A Discriminative Representation for Symbolic Image Similarity Evaluation, Workshop on Graphics Recognition, Brazil, [19] P. Leray, Réseaux Bayésiens : apprentissage et modélisation de systèmes complexes, novembre, [20]X.Li, Augmented naive bayesian classifiers for mixedmode data, December, [21] P.Naim, P.H.Wuillemin, P.Leray, O.Pourret, A.Becker. Réseaux bayésiens, Eyrolles, Paris, [22] E.G.M.Patrakis, Design and evaluation of spatial similarity approaches for image retrieval, In Image and Vision Computing Vol.20, pp.59-76, [23] L.Smail, Algorithmique pour les réseaux Bayésiens et leurs extensions, Thèse de doctorat, 2004 [24] J.E. Meng, W. Chen, W. Shiqian, High speed face recognition based on discrete cosine transform and RBF neural networks ; IEEE Transactions on Neural Networks, Vol. 16, No.3, pp , [25] M.Turk and A. Pentland. Eigenfaces for recognition. Journal of Cognitive Neuroscience, Vol.3, pp.71-86, [26] M.Jones and P.Viola, Face Recognition using Boosted Local Features, IEEE ICCV,2003. [27] A.S.Mian, M.Bennamoun, R.A.Owens, 2D and 3D Multimodal Hybrid Face Recognition, pp , [28] W.Zhao, R.Chellappa, P.J.Phillips, A.Rosenfeld, Face Recognition: A Literature Survey, ACM Computing Survey, pp , [29] J.Huang, B.Heisele, V.Blanz, Component-based Face Recognition with 3D Morphable Models,2003. [30] C.Garcia, G.Zikos, G.Tziritas, A Wavelet-based Framework for Face Recognition,Workshop on advances in facial image analysis and recognition technology, 5 th European conference on computer vision,1998. [31] M. Turk and A. Pentland, Face Recognition Using Eigenfaces, In IEEE Intl. Conf on Computer Vision and Pattern Recognition, pp , 1991 [32] M. Bartlett, J. Movellan, T. Sejnowski. Face Recognition by Independent Component Analysis. IEEE Trans. on Neural Networks, Vol.13,No.6,pp , [33] P. Belhumeur, J. Hespanha, D. Kriegman. Eigenfaces vs. Fisherfaces: Recognition Using Class Specific Linear Projection. IEEE Trans. on Pattern Analysis and Machine Intelligence, Vom.19, No.7, pp , [34] T. Ahonen, A. Hadid, M. Pietikâinen, Face Recognition With Local Binary Patterns, In European Conference on Computer Vision (ECCV), pp , [35] F. Cardinaux, C. Sanderson, S. Marcel, Comparison of MLP and GMM classifiers for face verification on XM2VTS, In 4th Intl. Conf. Audio- and Video-based Biometric Person Authentication, AVBPA, Vol.2688, pp , [36] S. Lucey and T. Chen, A GMM Parts Based Face Representation for Improved Verification through Relevance Adaptation, In IEEE Intl. Conf on Computer Vision and Pattern Recognition (CVPR), pp , [37] Y. Rodriguez and S. Marcel, Face Authentication Using Adapted Local Binary Pattern Histograms, In European Conference on Computer Vision (ECCV), pp , [38] L. Wiskott, J.M. Fellous, N. Kruger, C. Von Der Malsburg, Face Recognition By Elastic Bunch Graph Matching, In Intelligent Biometric Techniques in Fingerprint and Face Recognition, pp , [39] A. Nefian and M. Hayes, Hidden Markov Models for Face Recognition, In IEEE Intl. Conf. on Acoustics, Speech, and Signal Processing (ICASSP), Vol.5, pp , [40] F.Cardinaux, C.Sanderson, S. Bengio, User Authentication via Adapted Statistical Models of Face Images, IEEE Trans. on Signal Processing, Vol.54,pp , 2005 [41] R.Huang, V.Pavlovic, D.N.Metaxas, A Hybrid Face Recognition Method using Markov Random Fields, Proceedings of the Pattern Recognition, 17 th International Conference on (ICPR'04), Vol.3,2004. [42] S.Sharavanan and M.Azath, LDA based face recognition by using hidden Markov Model in current trends, International Journal of Engineering and Technology, Vol.1, pp.77-85, Khlifia Jayech is a Ph.D. student in the department of Computer Science of National Engineering School of Sousse. She obtained her master degree from Higher Institute of Applied Sciences and Technology of Sousse (ISSATS), in Her areas of research include Data Retrieval, Bayesian Network, Hidden Markov Model and Recurrent Neural Network, and Handwriting Recognition. Dr Mohamed Ali Mahjoub is an assistant professor in the department of computer science at the Preparatory Institute of Engineering of Monastir. His research interests lie in the areas of HMM, Bayesian Network, Pattern Recognition, and Data Retrieval. His main results have been published in international journals and conferences.

84 63 SbSAD: An Integrated Service-based Software Design Framework Mohamed Dbouk 1, Hamid Mcheick 2, Ihab Sbeity 1 1 Faculty of Sciences-I, Lebanese University Beirut, Lebanon 2 Dep. of Computer Science, University of Chicoutimi Quebec, Canada Abstract Phased software engineering process continues to be the most popular paradigm leading to devise and drawing-up all system architectural designs. In this paper we trying to explore and examine the most significant software engineering activity: Software architectural design. In this paper we discuss and evaluate an integrated service-based (the common and modern architectural styles upon which many systems are currently based) software architectural design framework called SbSAD. SbSAD is, mainly, built on top a proprietary micro-phased design process. In this paper, we first reconsider and refine such process in order to become more flexible. We, then, trying to evaluate this process by providing one devoted CASE-like prototype built using java technologies. Our approach consists of building overall software architectures while being based on the concept of business front-end services. The experiments show that: applying such strategy may cause some confliction with the so known SOA and may disorient both readers and designers. However, at the end, we testify that our service-based process should not have any direct connection with the SOA style. Working on some re-drawing and mapping rules leading to transcript SbSAD into SOA could characterize our future works. Keywords: Software architecture, Front-end services, SDLC, SOA, CASE tools, Data exchange, Dataflow. 1. Introduction Software architectural design and modeling persists and remains a crucial discipline, new and additional researches are reported each day. Researchers are incited and encouraged by the newest computer-related technologies and engineering policies. The most recent associated researches are concerned by topics like: ontological-based software architectural design, meta-modeling design and system distribution policies and strategies. Rather than the so known client/server software architectural style, SOA (Services Oriented Architecture) plays a factual and useful software engineering implantation strategy. However, by coming back to the most popular phased software development life cycle (SDLC), we observe that many efforts could also be deployed not only at software detailed design and modeling levels, but also at top software architectural design level. The challenge is, in fact: could new software engineering approaches benefit from accurate and fundamental concepts, techniques and technologies like software service-based global architectural design [11] and or like service-oriented architectural style. In other words, the software engineering experiments show that, in addition to spread out SDLC classical activities, an analytical concentration on software frontend services could perfectly help in producing an accurate software overall architecture. For this purpose, we started by concretizing such software architectural design philosophy with our service-based architectural design approach [11]. In this paper, we start by reviewing and reformulating the advocated above approach. We focus, then, on providing an empirical framework prototyping such approach. The framework consists of a CASE tool built on top of Java facilities and putting to gather: one devoted graphical user interface and one meta-modeled repository. By incorporating such meta-data, the tool forms an open and integrated CASE framework; towards multi-platforms code generator.

85 64 On the other hand, the experiments show that, when we approach the popular software element Service, an amalgam so arises between: the concept that we have introduced Service-based Architectural Design SbSAD and the so known SOA (Service-Oriented Architecture) style. Briefly, the SbSAD that we propose and we demonstrate/prototype seems as an integrated framework intended to help in software architectural design activities. Section 2 of this paper draws-up the associated concepts and basis. We, then, outline the related works. Principals that differentiate our approach from other approaches are also discussed in this section. Section 3 devises and reconsiders the foundations of our service-based approach. A devoted prototype (with experiments) is also outlined in this section. Section 4 tries to depict the differences between Service-based and Service-oriented terminologies. Finally, the conclusion and the future directions are outlined and drawn in section Background and Related works A computer system exists within an environment and has characteristics such as Boundaries and Front-End Interfaces. Building high-quality software is not an easy task; a wide range of software engineering paradigms have been proposed (e.g. object-orientation [3], design patterns [13] and software architectures [5]) either to make the engineering process easier or more flexible. System development methodologies evolved [14], we depict the following progression: SDLC - Systems Development Life Cycle [17], Structured Analysis and Design (SA&D) using data-flow diagrams, Data-Oriented Methodology using Entity-Relationship diagrams and Object-Oriented Methodology using UML facilities. Where the current trend is to use Object-Oriented Systems Analysis and Design, but many organizations are still using SA&D. Software design modeling techniques, that span stages in software lifecycle, are not standardized yet. The majority of modern software architectural strategies institute socalled decomposition methodologies. All these strategies employ some restricted vocabulary such as component (sub-system) and inter-components relationships. Goodness and robustness of the outlined software architectures are proportionally linked to the designers experience and maturity. Even so, computer system design is concerned about the overall structure of the system [4]. How is it broken into pieces? How do these pieces fit together? The best system design is one where the interaction between the subsystems is minimal. Data Management and designing end-user interfaces are vital. These steps involve and require robust software architecture and design knowledge. The experience of the design team, availability of predesigned components, capabilities of design automation tools, and the maturity of the process technology, all influence the degree to which an intended computer system must be decomposed. Good and modern designs do not ignore the past [18]. On the other hand, as reported in [16], ontology can be very useful in software engineering projects where development is focused not just on one application, but on a family of projects from the same domain. Ontology must be developed in a new taxonomy framework to describe the new concepts, properties, and relationships of the new project domains. In summary, the experience and maturity of system designers play a crucial and major role in system analysis and design process. An analysis by analogy could support most system decomposition activities. However, the decomposition process is typically conducted by designers based on their intuition and past experiences. The decomposition approach proposed by [20] tries to apply the clustering technique to support decomposition based on requirements and attributes. The approach supports the architectural design process by grouping closely related requirements to form a subsystem or module. Obtained decomposition and architectural styles or patterns are useful for developing a conceptual architecture as a representation of high-level design with critical components and connectors. On the other hand, [12][9] propose an approach using Ontology. The idea is to close the gap between requirements and components; they use semantic models a common language ODL (Ontology Design Language) for describing product requirements and component capabilities and constraints. Y. Cai and S. Huynh, From Drexel University in Philadelphia-USA, develop a Logic-Based Software Project Decomposition design representation called an Augmented Constraint Network (ACN) [6][7][8], they use the prototype tool Simon [1] to automatically

86 65 decompose a big ACN into a number of smaller sub- ACNs. Closely, P. Koopman proposed an elegant taxonomy of decomposition strategies [18]. The approach uses three attribute categories: Structures typically answer the question of "what", Behavior typically answers the questions of "how" and "when", and Goals as emergent design properties that satisfy the intended needs. To summarize, a common theme in this discussion and our approach is a fundamental coalition between requirements, attributes and clustering techniques. Many people have explored auto-clustering approaches to decompose an enormous dependent model into modules, such as Mancoridis s Bunch tool [21], which is based on heuristic fitness function. There are more works done in the same context as our approach, such as the feature-oriented research led by [2], [10] and [15]. The theoretical decomposition strategies, shown before, draw some interesting and useful decomposition methodologies when they refer to requirements. These strategies, also suffer from overcrowding, from the beginning, of the deployed information (attributes). Based on requirements and attributes, [20] applies the clustering technique to one huge and complex requirement/attribute matrix. The complexity is due to the early exploitation of information details! Otherwise, in spite of seniority and high abstraction level, the decomposition strategies drawn in [18] stay so elegant and plausible, but unfortunately, there is no pursuit. In addition to the above talk, the recent approach ([12] and [9]) using ontology represents an innovative direction using semantic models to describe both requirements and component. By contrast, the approach predicts a component specification, things that are not plausible in our case, because depicting components is a final goal for any design process. Lane [19] is similar to the logic-based approach [6][7][8] that models the structure of software systems as design spaces, they focus on functional choices. The logic-based approach works at abstract design level and applies formal modeling same as our approach. By contrast it applies an automatic analysis. Finally, the architectural design approach that we propose and reconsider defers from the above approaches by many things. We introduce the concept of software design contextual dimensions (business features): profiles, services, data and rules that should characterize any frontend services. We, especially, focus on software services because they, legally, represent the only visual and interactive software entry-points. We, also, consider nonatomic data; the experiments demonstrate that there are no real needs to know details about data from the beginning. Moreover, we continue to materialize the software engineering design scope of our approach by providing an emergent, integrated and open CASE-like framework SbSAD. 3. Service-based Design approach The approach, also called SbSAD (Service-based Software Architectural Design), which we are going to reconsider, occupies the first and crucial design stage in any traditional SDLC (fig. 1). This approach is a micro-phased process; it is constituted around five successive analytical and modeling microphases. The process takes, as input, a well written (requirements) document, and produces an Overall Software Architecture [11]. Engineering Requirements SbSAD Overall Software Architecture Traditional SDLC System Design Detailed Design Rest of the SDLC Fig. 1 SbSAD within a traditional SDLC The approach mainly focuses on computer business frontend services, the unique visible entry-points from any software system.

87 Formal and algebraic definitions This approach deals with three fundamental design features (called business/contextual dimensions) that characterize computer business services; business profiles, business data-items and business rules. Software engineering vocabulary - the process uses the following business features: - System (the target); computer or information system under design. - BService (BS); front-end business related software service, materializing an entry-point. - BProfile (BP); business domain materializing coherent set of computing activities. - BData (BD); data sets required by the computing activities. - BRule (BR): implicit and/or explicit business and constitutional rule, pre or post depicted against the engineering activities. i BS j Rule 5 BP 1 i BS 1 ij BD 1 Computer System Rule 6 BPi i BS j BD i j ij BD k Composition Rules: 1, 2, 3 Sharing Rules: 4, 5 Services grouping Rules: 6 Rule 1 BPn Rule 2 Rule 3 Fig. 2 Computer system structural form in SbSAD i BS n BModule ij BD n Basically, the SbSAD process leans, indeed, on some basic constitutional architectural rules. These rules are qualified as algebraic: Structural rules: Rule 1- Rule 2- Rule 3- contains System BP,..., BP,..., BP (1) i contains i 1 1 i j i n i n ij n BP BS,..., BS,..., BS (2) i j contains ij 1 BS BD,..., BD,..., BD (3) Engineering rule: Rule 4- BProfiles may share BServices, formally: i j i j BS, BS where BS BS (4) BPi.BS BP j.bs Benefits: Engineering reusability of services Behavioral rule: Rule 5- BProfiles may share BData. Sharing issue may be direct or indirect, by similarity (same data) or by aggregation (ETL like method), formally: BD, BD i j where BD BP i.bs.bd BPj.BS.BD Benefits: Business workflow depiction. i i ij k BD Extra-Structural rule: Rule 6. BServices for one BP may be regrouped by sub-bprofiles in order to form the so called Business Modules BM (BModules). Benefits: incremental system engineering building The SbSAD functional process By revising the process, we observe, that there is no real need, to suddenly depict SbSAD s huge amount of conceptual features from the beginning. Instead, it will be more efficient to proceed incrementally, feature by feature. The experiment indeed, shows that the micro-phases of this process may be applied freely; activities must be reiterated until exhausting all features. The experiment also shows that the identified serviceclusters (subsystems/components) would be reexamined in terms of modules; a module regroups one named homogenous set of front-end services and shares the same data-items with other modules inside one named business profile (Services cluster). The system requirements R represent the main source of information; the following engineering design activities might stimulate the above process: a. read/analyze R, depict/identify one or more Business Profiles BP. j j (5)

88 67 b. read/analyze R, depict/identify one or more Business services BS. c. read/analyze R, depict/identify one or more Business Data BD. d. explore BS, attach the BS to BP. e. explore BD, attach the BD to BS. f. read/analyze R, expand BD, depict/identify BDexp g. explore BP/BD, BDexp, depict/identify BP interrelationships, associate BProfiles mutually. h. explore BP/BS, depict/identify Business modules BM, de-attach BS from BP, attach BS to BM and BM to BP. System Requirements B. Profile depiction Business Profiles B. Module depiction Business Modules System Architecture building B. Service depiction Business Services B.Data depiction Business Data B.Data Expanding Business Data details (BDexp) Dataflow depiction BP Interrelationships System Overall Architecture Process workflow Process reiterations Features articulation and refinement Fig. 3 SbSAD process pictogram If we examine the above activities, we can predict that the order is insignificant, most of them may be permuted; swaps are permitted with respect to the dependability criteria, they could be performed individually. For example, designer may perform a (non complete) sequence like this: a, b, d, c, g, e or this: b, a, d, c, b, e, g, etc. We observe that activity numbered d depends on a and b, etc. However, the best way to represent such engineering design activities is by using an overall pictogram (fig. 3). Designers can, then, build the overall architecture incrementally same as the case if we use an appropriate editing-tool (software overall architectural builder). At the end, the required SbSAD s features should be situated and totally explored. Designers, charged to elaborate a system design, start by reading the requirements document. They depict conceptual features one by one. Each time they identify one feature (business profile, and/or business service as well as business data), they could articulate/associate it to the adequate partner (structural rules). Designers could refer to the requirements document each time they try to perform one design activity (ovals in fig. 3). Finally, the above pictogram materializes, transparently, all SbSAD predicted architectural rules, and produces, at the end, the intended architectural structure/design (fig. 2) Prototyping and validation Practically, SbSAB is in the course of prototyping. The validation of the above functional process is divided over tree stages: - a GUI (as a CASE tool) materializing the operational process, - building the related and required metadata, - and validation via real use case. The intended tool tends to integrate engineering technologies such as platforms related code generation. SbSAD as a CASE tool: The issue is to provide a user friendly (fig. 4) graphical user interface materializing the different system design engineering functionalities. The question now, is: How does such tool operate? A brief abstraction of the operational process is given in fig. 5. However, the main purpose of this tool is to provide a useful interactive framework building and producing the overall system architecture.

89 68 Main control area especially, includes the overall system structural and dataflow behavioral features. Treelike Project Manager Area Workspace - main area (System-Design editor) Zooming area: Project s related SbSAD-features SbSAD-System BProfile * Alerts area Fig. 4 The SbSAD s interactive graphical interface DataLink 0..* 0..* BModule 0..* 1..* 1..* BService 0..* The tool incorporates one crucial (integrated) piece that supports the implementation software engineering phase; it refers also to one devoted back-end meta-model. 0..* BDataExp 0..* 0..1 BData 0..* Open Project Repository Meta data Platform s Meta data System design editing Design verification Integrated peace Destination Platform Architecture code-generation Generated Code System Architecture Editor Fig. 5 SbSAD s CASE tool operational process Back-end Meta model: The SbSAD s operational process refs to some devoted meta-data (fig. 6), a crucial piece that consists of a collection of back-end related data (enclosed class diagram). However, the data model incorporates all SbSAD related features and concepts. In addition to information cataloging the system required data, the data-model, Fig. 6 Repository UML class diagram Finally, this information forms an editable repository that could be strongly used by the tool. Use case and validation: The first and empirical version of SbSAD is now operational; it is built on top of Java technology. Many improvements and enhancements are planed. The enclosed figure (fig. 7) draws a typical case: Sale/Purchase-Management Information System (S/P- MIS). The given sample illustrates and shows the following (SbSAD) features: - Business Profiles (SubSystems): Stock-Manager (StkMg), Point-Of-Sale (PoS), OLAP-Manager (OlapMg), etc. - StkMg may incorporate: Produts-Nomination (PdNames) and Inventories (InvMg) etc. as modules. - Intiate-PoS and Close-PoS are data-links relating StkMg to PoS considered subsystems and vice versa. Those links could be infected by the chosen strategy to materialize the data sharing issue; online or offline (differed) mapping. A crucial data link (ETL-like) should exist between StkMg and OlapMg. Finally, many kinds of data items could characterize the S/P-MIS ; products, commands, customers (if considered), suppliers, cashers, etc.

90 69 Fig. 7 The SbSAD s prototype; integral graphical user interface view. 4. SbSAD vs. SOA As mentioned before, the SbSAD process is concerned by providing the overall system/software architecture. The process represents one of two major design phases (figures 1 & 8). At this stage, the designer doesn t have to worry about the manner, according to which the target system could be implemented and deployed. Such task, strictly, comes after. The, so known, SOA (Service Oriented Architecture) policy and strategy is, practically, especially concerned with the implementation and system deployment issues. Engineering Requirements System Design SbSAD Detailed Design Devoted SbSAD-SOA SDLC SbSAD / SOA mapping space Implementation So, the question that arises is: How could we qualify the connection/relationship between SbSAD and SOA? First of all, the SbSAB approach is intended to be a design process while SOA is seen as an implementation related architectural style (fig. 8). SOA is here Rest of the SDLC Fig. 8 SbSAD / SOA Space

91 70 Practically, a close mapping of SbSAD s outcomes into SOA methods and techniques should produce one devoted and powerful software design and implementation process. Such mapping process could include one dedicated set of rules that drawing out one regular SbSAD-SOA software engineering policy. 5. Conclusion and Future work We started this work by establishing the main software architectural design features, context and background. We devised, in this paper, one dedicated software engineering framework (SbSAD), it consists of an integrated CASE-like tool. The tool is prototyped using Java technology, the first version gives good results, additional features and improvements are planed. However, we, in this paper, reconsidered and reevaluated the service-based software architectural design process. We talked, especially, about the formal definition as well as the operational modalities of the process. In addition to the above talk, the experiments show, that the process might, easily, be extended. It could incorporate new design facilities; detailed design, multi-platform related design, etc. To conclude, as future work, we plan to work on the mapping issues between SbSAD outcomes and SOA style. References [1] Baldwin, C.Y. and Clark, K.B., Design Rules, Vol. 1: The Power of Modularity. Publisher: The MIT Press, , 483 Pages, ISBN: [2] Batory, D., Singhal, V., Thomas, J., Dasari, S., Geraci, B. and Sirkin, M., The genvoca model of software-system generators, IEEE Software, 11(5):89 94, Sept [3] Booch, G., Object-oriented analysis and design with applications, 1994, Addison Wesley. [4] Bruegge, B. and Dutoit, A.H., Object Oriented Software Engineering Using UML, Patterns and Java", Second Edition. Pearson Education International, 2004 [5] Buschmann, F., Meunier R., Rohnert, H., Sommerlad, P. and Stahl, M., A System of Patterns, 1998, Wiley. [6] Cai, Y., Modularity in Design: Formal Modeling and Automated Analysis. PhD thesis, Univ. of Virginia [7] Cai, Y. and Sullivan, K., Modularity analysis of logical design models. In 21th IEEE/ACM Int. Conf. on Automated Software Engineering, Tokyo, JAPAN, [8] Cai, Y. and Simon, K.S., A tool for logical design space modeling and analysis. In 20th IEEE/ACM Inter. Conf. on Automated Software Engineering, Long Beach, California, USA, Nov [9] Cardei, I., "An Approach for Component-based Design Automation", Whitepaper, Florida Atlantic University press 2006 [10] Czarnecki, K. and Eisenecker, U., Generative Programming: Methods, Tools, and Applications. Addison-Wesley Professional, 1st edition, Jun [11] Dbouk, M., Sbeity, I. and Mcheik, H., Towards Service- Based Approach; Building Huge Software Architectural Designs, IJCNDS, 2011, v6 (forthcoming). [12] Fonoage, M., Cardei, I. and Shankar, R, "Mechanisms for Requirements Driven Component Selection and Design Automation" the 3rd IEEE Systems Conference, Vancouver, Canada, [13] Gamma, E., Helm, R., Johnson, R. and Vlissides, J., Design Patterns, 1995, Addison Wesley. [14] George, J.F., Batra, D. and Valacich, J.S., Object- Oriented Systems Analysis and Design, ISBN: , ISBN-13: , Published by Prentice Hall, 2006 [15] Goguen, J.A., Reusing and interconnecting software components. IEEE Computer, 19(2):16 28, Feb [16] Hesse, W., Ontologies in the software engineering process. In R. Lenz et al., editor, EAI 2005 Proceedings of the Workshop on Enterprise Application Integration [17] Hoffer, J.A., George, J.F. and Valacich, J.S., Modern Systems Analysis and Design (5th Edition), ISBN-10: , ISBN-13: , Published by Prentice Hall, 2008 [18] Koopman, P., "A taxonomy of decomposition strategies based on structures, behaviors, and goals", 1995 Conference on Design Theory and Methodology, Boston, September [19] Lane, T.G., Studying software architecture through design spaces and rules, Technical Report CMU/SEI-90-TR-18, CMU, [20] Lung, C. and Zaman, X. M., "Software Architecture Decomposition Using Attributes" Carleton University, Ottawa, Ontario press, Canada [21] Mancoridis, S., Mitchell, B., Rorres, C., Chen, Y. and Gansner, E., Using automatic clustering to produce highlevel system organizations of source code. In Proceedings of the 6th Inter. Workshop on Program Comprehension (IWPC 98), pp , June Mohamed DBOUK, received a Bachelor s Honor in Applied Mathematics; Computer Science, Lebanese University, Faculty of Sciences, and a PhD from Paris-Sud 11 University (Orsay-France), He is a full time Associate Professor, at the Lebanese university - Faculty of Sciences-I, Dep. of Computer Science. His was ( ) the director of this Faculty, and he is the Founder and Coordinator of the research master M2R-SI: Information System. His research interests include Software engineering, Information systems, GIS, Cooperative and Multi-Agent Systems, Groupware. He participates in many international projects. Hamid Mcheick is currently an associate professor in Computer science department at the University of Quebec At Chicoutimi (UQAC), Canada. He holds a master degree and PhD. in software engineering from Montreal University, Canada. Ihab Sbeity occupies a full time position in Computer Science Department at the Lebanese University. He holds a PhD.in performance evaluation and system design from Institut National Polytechnique de Grenoble, France in 2006

92 71 Success Rules of OSS Projects using Datamining 3-Itemset Association Rule Andi Wahju Rahardjo Emanuel 1, Retantyo Wardoyo 2, Jazi Eko Istiyanto 3, Khabib Mustofa 4 1 Bachelor Informatics, Faculty of Information Technology, Maranatha Christian University Bandung, West Java, Indonesia 2, 3, 4 Department of Computer Science and Electronics, Gadjah Mada University Yogyakarta, Indonesia Abstract We present a research to find the success rules of 134,549 Open Source Software (OSS) Projects at Sourceforge portal using Datamining 3-Itemset Association Rule. Seventeen types of OSS Project's data are collected, classified, and then analyzed using Weka datamining tool. The Datamining 3-Itemset Association Rule is used to find the success rules of these projects by assuming that the success of these projects are reflected by the number of downloads. The result are formulated into 9 success rules that may be used as guidelines by future initiators of OSS Project and other developers to increase the possibility of success of their projects. Keywords: Open Source Software Project, Datamining Association Rule, Success Rule, sourceforge.net 1. Introduction Open Source Software (OSS) is one of the current trends in Information Technology, especially in the field of Software Engineering. Once thought only as the sharing playground for researchers, academics and programmer enthusiasts during their spare time, this methodology is evolved into one of the mainstream software development methodology challenging the already established software engineering disciplines. Some success stories about this OSS Projects such as Apache Web Server, Linux Operating System, Openoffice.org productivity suite, Mozilla Web Browser, and many more. Despite the apparent success stories relating to OSS projects, there are many more projects using this scheme which are failed. Some approaches or guidelines need to be discovered to assist an initiators and contributors of OSS Projects in increasing the chance of success for the project. We believe that these approaches / guidelines could be found by studying the existing small to medium sized OSS Projects to find their success rules. In our previous research by gathering OSS Project's information from Sourceforge portal and using Datamining 2-Itemset Association Rule already found 6 success factors [5]. This research is further exploration from this research in which we are using 3-Itemset Association Rule to find additional or more specific success rules. This paper is organized as follows: Section 2 describes the current studies on OSS Project s success factors, Section 3 describes the theoretical background of OSS Projects and Datamining Association Rule. The Datamining processes are described in Section 4 with the interpretation of the result into the OSS success factors is shown in Section 5. The conclusion is described in Section Current Studies on OSS Project Success Many studies have been conducted to identify the key success factors of OSS Projects. One approach of the study is by studying the processes of many large and successful projects, such as the study on Debian GNU/Linux [12], FreeBSD [4], Apache Web Server [10], OpenBSD [7], Apache against Mozilla [9], Arla against Mozilla [2], and some 15 popular OSS Projects [8]. This approach may provide excellent examples about how large OSS project works; however, these large and successful OSS Projects already have established process and organization involving large many developers and other stakeholders that are difficult to be implemented by small and medium sized OSS Projects. The study of small and medium sized OSS Projects that considered successful are more relevant compared to the study on large and mature OSS Projects since all of these projects are usually start from small size. In our previous research, by gathering OSS Project's information from Sourceforge portal by using Datamining 2-Itemset Association Rule, we have found 6 success factors [5]. Further elaboration used in this research is by

93 72 using 3-Itemset Association Rule to find more detail or additional success rules that contribute to the success of OSS Projects. The subject of the research is still the small to medium sized OSS Projects hosted in one of the most popular web portal which is Sourceforge. At the time of this research (January, 2010), this portal had 160,141 registered projects; and in this research, 134,549 OSS Projects are selected and their data are extracted and analyzed using Datamining 3-Itemset Association Rule. 3. Theoretical Background 3.1 Open Source Software Projects Open Source is a software development methodology based on several distinct characteristics: The source code of the application is freely available for everybody to download, improve and modify [11]. People who contribute to the development of the Open Source projects is forming a group called Open Source Community which is voluntary [3]. The primary concern of the developers in Open Source Software Projects are building features and fixing bugs [6]. In order to develop software application in OSS Project, a project initiator may use the service from OSS Development portal such as sourceforge.net, launchpad.net, Google code, etc. The sourceforge.net portal is chosen since it covers more than 70% of total OSS Projects from these popular portals [5]. 3.2 Datamining Association Rule Datamining is a technique to find hidden structure and relationship in a large number of population [1]. The knowledge about these structure and relationship is discovered by using two methods which are predictive (predicting unknown value or future value of a variable), and descriptive (finding human readable patterns). In this research, the descriptive method is selected in this research since it is intended to find the human readable patterns from all the data being collected from Sourceforge.net. In this method, there are several techniques / rules that may be chosen which are Classification, Segmentation / Clustering, Association, etc. The Association Rule is selected since it is able show the dependency between one parameter to another parameter of some large collections of data. The Association Rule will find dependency rule which will predict the appearance of an itemset (Consequent) based on the appearance of other itemset / itemsets (Antecedent). The 3-Itemset Association Rule has two Antecedents connected with logical AND, and a single Consequent. The 3 Itemset may be stated as: {X, Y} => {Z}...(1) Where X is the first Antecedent (Antecedent1) and Y is the second Antecedent (Antecedent2), and Z is the Consequent. X and Y must appear at the same time as the cause and Z is the result with some certainty values called Support and Confidence. The value of Support ({(X,Y),Z}) shows the number of transaction containing item X AND Y and item Z against total population, whereas the value of Confidence((X,Y)=>Z) shows the probability of the occurrence of item Z if a transaction containing item (X,Y) and Z. In this research, the rules are of in the interest if it has the minimum Support of 10% and the minimum Confidence of 50%. The value of 10% for the Support is selected since it will represent significant proportion of the entire population, and the value of 50% in Confidence is selected since it also represent half or more in terms of probability of occurrence. 4. Datamining Processes Table 1 shows the data description of the OSS project recorded from sourceforge.net portal. There are 17 types of data parameters being recorded for each of the OSS Projects. Table 1: OSS Project Data Description No Parameter Type Remark 1 Name Text The name of the project 2 Audience Text Intended audience 3 Database Text Database environment used 4 Description Text Project description 5 Developer Text User name of developer 6 Development Status Text Status of project development 7 Download Integer Number of download 8 Filename Text Name of downloadable file from project's front page 9 File Size Text Size of downloadable filename 10 License Text Applicable license 11 Operating Text Applicable operating system

94 73 No Parameter Type Remark System 12 Programming Language Text Programming language used 13 Review Text Review from user 14 Thumb Integer Recorded thumb up and thumb down from user 15 Topic Text Applicable topic for the project 16 Translation Text Available language translation 17 User Interface Text Applicable user interface The parameters such as audience, database, developer, development status, license, operating system, programming language, review, topic, translation, and user interface are having zero to many parameters. The count of these parameters is also considered as the parameters used during datamining process. The total parameters being recorded are more than 27 parameters if it includes the count of these parameters. 4.1 Data Collection Process The data collection process of OSS Projects from Sourceforge was conducted by creating custom-made PHP script crawler. The collecting process was conducted in three phases: Recording summary of projects (from link to record most of the parameters shown in table 1. Recording more detail information by crawling each individual project link page to record the developer, project description, filename, file size, number of thumbs (up and down), and reviews. Filling the missing information, finding and deleting duplicates and then generating count data from multiple-value parameters. The collection process was taking about 9 weeks to complete starting early January 2010 through the end of February Out of 160,141 OSS Projects registered from the portal, the crawler was able to collect data from 134,549 unique projects stored in 27 tables with total 3,115,085 records. 4.2 Data Classification Process The next process was the classification of the data. Most of the parameters need to be classified in some categories with enough number of population in order to gain meaningful rules. Audience: There are 121,095 OSS Projects (90%) of the recorded projects that list the audiences of its software project. There more than 23 distinct values of project's audience which are then classified into three classes which is Specific Audience (42.54%), Developers (29.10%), and Common Users (28.36%). Database Environment: There are only 30,335 OSS Projects (22.55%) of recorded projects are using at least one database. The OSS Project with database are classified as MySQL (31.25%), SQL-based (27.17%), API-based (20.48%), Text-based (15.42%), and Other (5.68%). Project Description: Each OSS Project has a description to state the purpose of project. The project description is mostly short sentence / paragraph with the peak at about 36 words. The project description is classified into three categories which are short (< 26 words 43.78%), middle (26-36 words 32.58%) and long (> 36 words 23.64%). Development Status: There are 128,215 OSS Projects (95.29%) of recorded projects that list the development status of their projects. The classification is based on the development status of the project which are 1 Planning (18.84%), 2 Pre-Alpha (15.15%), 3 Alpha (17.15%), 4 Beta (24.05%), 5 Production / Stable (20.56%), 6 Mature (1.83%), and 7 Inactive (2.42%). Number of Download: Table 2 shows the statistics of the number of downloads of OSS projects. The number of download is 0 may means that there are no download or the project does not have any downloadable file. Table 2. Statistics about Number of Download Download Population Percentage 0 / NA 55, % , % , % , % 10,000-99, % 100, , % 1,000,000-9,999, % 10,000, % Note: NA - not available (downloadable file is not yet available). The number of download of OSS projects in categorized as none (41.61%), hundred or less (32.42%), and

95 74 thousands or more (25.97%). In this research, the number of download is assumed as the indication of success in OSS Project. If an OSS Project is successful, it will be accepted by many users that is indicated by the large number of downloads for the project. Therefore, the number of download is classified into three categories which are none (0 / NA download), hundreds or less (1 up to 999 downloads), and thousands or more (more than 1000 downloads). The Association Rule that has Download Thousands or more as Consequent with any possible combinations of two other Antecedents are the interested rules. Filename: This experiment only record the filename and its size listed on the projects' site on the first page ( This filename is not necessary the only available filename, and there is also no guarantee that the filename is always the source code of the project. The filename is then classified based on its extension which are zip (47.27%), tar.gz (29.72%), jar (7.42%), tar.bz2 (5.82%), tgz (5.02%), rar (2.46%), and other format (2.28%). File Size: The size of the downloadable filename was also recorded and then categorized based on its order of magnitude (BYTES, KB, MB, or GB). The classifications are BYTES (0.26%), KB (71.90%), MB (27.79%), and GB (0.06%). License: There are 131,777 OSS Projects (97.94%) of recorded projects that list the applicable license for the project. There are 75 distinct values for the license for OSS projects, and they are classified into GPL (61.57%), LGPL (10.64%), BSD License (6.75%), Apache License (3.78%), Public License (3.40%), MIT License (2.62%), AFL (2.62%), Mozilla License (1.40%), and Other (8.28%). Operating System: There are 111,760 OSS Projects (83.06%) of recorded projects that list the applicable Operating System. There are 85 distinct Operating System for the OSS projects which are then classified as Linuxlike (35.63%), Windows (34.25%), Cross-Platform (23.19%), or Other (6.93%). Programming Language: There are 127,247 OSS Projects (94.57%) of recorded projects that list the applicable programming language. There are 97 distinct programming languages for the OSS projects which are then classified into Java (20.10%), C++ (16.27%), Other OOP (7.45%), C (14.91%), PHP (13.14%), Other Script-based (18.57%), or Other (9.56%). Thumb (Up and Down): There are only 16,829 OSS Projects (12.50%) of recorded projects that being thumbreviewed (users give either thumb up or thumb down). The classifications of thumb are single (48.58%), two or three (24.34%), four to ten (15.86%), and eleven or more (11.21%). Topic: There are 440 distinct topics for the OSS Projects which are then classified into 6 categories which are Software Development (19.97%), Internet/Networking (17.41%), Data Management (17.37%), Games/Entertainment (14.49%), Scientific/Engineering (11.65%), Other topic (19.10%). Translation: There are 77,269 OSS Projects (57.43%) of recorded projects that list the available language translation. There are 67 distinct values for available language translation that is then classified into three classes which are English (59.18%), European (33.85%), and Other (6.98%). User Interface: There are 97,302 OSS Projects (72.32%) of recorded projects that list the available user interface for the project. There are 60 distinct values which is then classified into 4 classes which are Desktop-based (46.91%), Web-based (25.57%), Text-based (17.13%) and Other (10.39%). Parameter's Count: The count of parameters are also recorded and classified. The classification is categorized into only three classes which are one, two, and three or more. The parameters that are classified in this scheme are audience count, database count, developer count, development status count, license count, operating system count, programming language count, review, and user interface count. 4.3 Result of Datamining 3-Itemset Association Rule The process of Datamining Association Rule was conducted with 3-Itemset. There are 277 possible combinations of 3-Itemset that have been processed using Weka resulting in 111 interesting rules that surpass the minimum values of Support and Confidence. The result of Datamining Association rule that have Download - Thousands or more as Consequent with Confidence more than 50% and Support more than 10% are the interested rule. Due to the limited table space, the value of Antecedent1 and Consequent ( Download Thousands or more ) are not stated in the tables. Audience: Table 3 shows the result with Audience Common Users as Antecedent1.

96 75 Table 3 Result for Antecedent1: Audience-Common Users Antecedent2 Analysis Parameter Class Pop. Sup. Conf. License GPL % 51.84% User Interface Desktop-based % 58.37% Audience Count: Table 4 shows the result with Audience Count One as Antecedent1. Table 4. Result for Antecedent1: Audience Count One Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Review Count three or more % 93.75% Review Count one % 69.33% Total Thumb single % 62.50% User Interface Desktop-based % 51.10% Database: Table 5 shows the result with Database SQL-based as Antecedent 1. Table 5. Result for Antecedent1: Database SQL-based Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Review Count one % 74.06% Review Count three or more % 94.70% Database Count: Table 6 shows the result with Database Count one as Antecedent1. Table 6. Result for Antecedent1: Database Count One Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Development Status 5 Production / Stable % 55.57% Review Count one % 63.65% Review Count three or more % 90.61% Total Thumb single % 54.53% Translation European % 63.29% Developer Count: Table 7 shows the result with Developer Count one as Antecedent1. Table 7. Result for Antecedent1: Developer Count One Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Review Count three or more % 92.29% Review Count one % 66.50% Total Thumb single % 60.41% User Interface Desktop-based % 50.32% Development Status: Table 8 shows the result with Development Status 5 - Production/Stable as Antecedent1. Table 8. Result for Antecedent1: Development Status 5 Production/Stable Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Database Count one % 55.57% License Count one % 59.68% Programming Language Count one % 57.45% Review Count one % 80.29% Review Count three or more % 97.26% Translation European % 80.31% Translation English % 67.24% Development Status Count: Table 9 shows the result with Development Status Count one as Antecedent1. Table 9. Result for Antecedent1: Development Status Count One Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Operating System Linux-like % 51.54% Operating System Windows % 54.38% Review Count one % 70.81% Total Thumb single % 63.91% Translation English % 51.00% User Interface Desktop-based % 53.52% Filename: Table 10 shows the result with Filename zip as Antecent1. Table 10. Result for Antecedent1: Filename Zip Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Review Count three or more % 94.12% Review Count one % 73.28%

97 76 Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Total Thumb single % 65.33% Translation English % 55.87% Translation Count one % 51.84% User Interface Desktop-based % 56.79% File Size: There are two groups of result which are either Size KB or Size MB as Antecedent1. Table 11 shows the result with Size KB as Antecedent1. Table 11 Result for Antecedent1: Size KB Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Operating System Linux-like % 50.12% Operating System Windows % 54.04% Review Count one % % Total Thumb single % 64.15% User Interface Desktopbased % 51.92% Table 12 shows the result with Size MB as Antecedent1 Table 12. Result for Antecedent1: Size MB Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Review Count one % 77.06% Review Count three or more % 96.04% License: Table 13 shows the result with License GPL as Antecedent1. Table 13. Result for Antecedent1: License GPL Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Audience Common Users % 51.84% Review Count three or more % 94.26% Review Count one % 70.76% Total Thumb single % 63.35% License Count: Table 14 shows the result with License Count One as Antecedent1. Table 14. Result for Antecedent1: License Count - One Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Development Status 5 Production / Stable % 59.68% Operating System Windows % 54.74% Operating System Linux-like % 51.81% Review Count three or more % 94.17% Review Count one % 71.51% Total Thumb two to three % 75.87% Total Thumb single % 64.26% Translation English % 51.31% User Interface Desktop-based % 53.93% Operating System: There are two groups of result which are either Operating System Linux-like or Operating System windows as Antecedent1. Table 15 shows the result with Operating System Linux-like as Antecedent1. Table 15. Result for Antecedent1: Operating System Linux-like Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Development Status one % 51.54% Count License Count one % 51.81% Size KB % 50.12% Programming Language Count one % 50.31% Total Thumb single % 74.75% Translation Count one % 52.77% Table 16 shows the result with Operating System Windows as Antecedent1. Table 16. Result for Antecedent1: Operating System Windows Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Development Status one Count % 54.38% License Count one % 54.74% Size KB % 54.04% Programming Language Count one % 52.57% Review Count three or more % 96.01% Total Thumb single % 70.49% User Interface Count one % 54.22%

98 77 Operating System Count: Table 17 shows the result with Operating System Count one as Antecedent1. Table 17. Result for Antecedent1: Operating System Count: One Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Review Count one % 69.36% Review Count three or more % 93.26% Total Thumb single % 63.34% Total Thumb two to three % 75.38% Programming Language Count: Table 18 shows the result with Programming Language Count One as Antecedent1. Table 18. Result for Antecedent1: Programming Language Count - One Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Development Status 5 Production % / Stable 57.45% Operating System Windows % 52.57% Operating System Linux-like % 50.31% Review Count one % 69.42% Review Count three or more % 93.43% Total Thumb single % 63.27% Translation European % 61.22% User Interface Desktop-based % 51.67% Review Count: There are two groups of result which are either Review Count one or Review Count three or more as Antecedent1. Table 19 shows the result with Review Count One as Antecedent1. Table 19. Result for Antecedent1 : Review Count One Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Audience Count one % 69.33% Database SQL-based % 74.06% Database Count one % 63.65% Developer Count one % 66.50% Development Status 5 Production / Stable % 80.29% Development Status one Count % 70.81% Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Filename zip % 73.28% License GPL % 70.76% License Count one % 71.51% Operating System one % 69.36% Count Programming one % 69.42% Language Count Size MB % 77.06% Size KB % % Translation European % 82.49% Translation English % 79.44% Translation Count one % 75.47% User Interface Desktop-based % 79.42% User Interface Count one % 70.52% Table 20 shows the result with Review Count three or more as Antecedent1. Table 20. Result for Antecedent1: Review Count three or more Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Audience Count one % 93.75% Database SQL-based % 94.70% Database Count one % 90.61% Developer Count one % 92.29% Development Status 5 Production / Stable % 97.26% Filename zip % 94.12% License GPL % 94.26% License Count one % 94.17% Operating System Windows % 96.01% Operating System one % 93.26% Count Programming one % 93.43% Language Count Size MB % 96.04% Translation European % 98.67% Translation English % 95.61% Translation Count one % 94.49% User Interface Desktop-based % 96.30% User Count Interface one % 93.70%

99 78 Thumb: There are four groups of result which are Thumb single, Thumb two or more, Thumb four to ten, or Thumb eleven or more as Antecedent1. Table 21 shows the result with Thumb single as Antecedent1. Table 21. Result for Antecedent1: Thumb - Single Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Topic Count one % 59.96% Audience Count one % 62.50% Database Count one % 54.53% Developer Count one % 60.41% Development Status one % 63.91% Count Filename zip % 65.33% License GPL % 63.35% License Count one % 64.26% Operating System Linux-like % 74.75% Operating System Windows % 70.49% Operating System one % 63.34% Count Programming one % 63.27% Language Count Size KB % 64.15% Translation Count one % 69.88% User Interface Desktopbased % 71.59% User Interface Count one % 63.71% Table 22 shows the result with Thumb two to three as Antecedent1. Table 22. Result for Antecedent1: Thumb Two or Three Antecedent2 Analysis Parameter Class Pop. Sup. Conf. License Count one % 75.87% Operating System one Count % 75.38% Translation Count one % 80.86% User Interface Desktop-based % 80.99% Table 23 shows the result with Thumb four to ten as Antecedent1. Table 23. Result for Antecedent1: Thumb Four to ten Antecedent2 Analysis Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Translation Count one % 91.71% Table 24 shows the result with Thumb eleven or more as Antecedent1. Table 24. Result for Antecedent1: Thumb Eleven or More Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Translation European % 99.75% User Interface Desktop-based % 99.52% Translation: There are two groups of result which are either Translation English or Translation European as Antecedent1. Table 25 shows the result with Translation English as Antecedent1. Table 25. Result for Antecedent1: Translation - English Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Development Status 5 Production % 67.24% / Stable Development Status one % 51.00% Count Filename zip % 55.87% License Count one % 51.31% Review Count three or more % 95.61% Review Count one % 79.44% Translation Count one % 50.19% User Interface Count one % 50.66% Table 26 shows the result with Translation European as Antecedent1. Table 26 Result for Antecedent1: Translation - European Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Database Count one % 63.29% Development Status 5 Production % 80.31% / Stable Programming one % 61.22% Language Count Review Count three or more % 98.67% Review Count one % 82.49% Total Thumb eleven or more % 99.75%

100 79 Translation Count: Table 27 shows the result with Translation Count one as Antecedent1. Table 27. Result for Antecedent1: Translation Count - One Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Filename zip % 51.84% Operating System Linux-like % 52.77% Review Count one % 75.47% Review Count three or more % 94.49% Total Thumb single % 69.88% Total Thumb two to three % 80.86% Total Thumb four to ten % 91.71% Translation English % 50.19% User Interface Desktop-based % 54.78% User Interface: Table 28 shows the result with User Interface Desktop-based as Antecedent1. Table 28. Result for Antecedent1: User Interface Desktop-based Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Audience Common Users % 58.37% Audience Count one % 51.10% Developer Count one % 50.32% Development Status one % 53.52% Count Filename zip % 56.79% License Count one % 53.93% Programming Language Count one % 51.67% Review Count three or more % 96.30% Review Count one % 79.42% Size KB % 51.92% Total Thumb single % 71.59% Total Thumb two to three % 80.99% Total Thumb eleven or more % 99.52% Translation Count one % 54.78% User Interface Count: Table 29 shows the result with User Interface Count One as Antecedent1. Table 29. Result for Antecedent1: User Interface Count One Antecedent2 Analysis Parameter Class Pop. Sup. Conf. Operating System Windows % 54.22% Review Count one % 70.52% Review Count three or more % 93.70% Total Thumb single % 63.71% Translation English % 50.66% 5. Interpreting the Result Combining the interpretation from Datamining 3-Itemset Association Rules are the success factors that should be followed by the project initiators and other developers to increase the probability of success of their OSS Projects. The interpretation is done qualitatively by noticing the frequency of appearance of Antecedent1 and Antecedent2 of table 3 through table 29. These success rules are: 1. Project should target for common users as audience. 2. Project source code should already in 5 Production / Stable development status. 3. Project should work on either Linux-like or Windows operating system. 4. Project should be reviewed and thumb-reviewed by at least one users. Project with windows operating system should have at least three reviews. 5. Project should have Desktop-based User Interface. 6. Project should select a single type of license, preferable GPL license. 7. Project has filename in zip format with size in either KB or MB in magnitude. For project with file size MB, it needs three or more reviews. 8. If the project is using database environment, select SQL-based database, and it should be reviewed by at least one user. 9. Project should have either English or European language translation. Rule number 1, 2, 3 and 9 are similar to the previous findings using 2-Itemset Association Rule [5], rule number 4 and 7 are more specific compared to the previous findings, and rule number 5, 6, and 8 are new rules. It is also observed that some freedom is still available for project initiator to decide such as the topic, programming language and description of his/her project without affecting the number of download. Some caution should be considered regarding to these rules. The subject being researched is small to medium OSS Projects from Sourceforge that may not reflect the whole population of OSS Projects that are small, medium

101 80 and large scale using OSS development portals or hosting in their own website. These result should also verified using OSS Project data from other portal such as launchpad.net, Google code, etc. to verify their validity. 6. Conclusions We present the Datamining 3-Itemset Association Rule of 134,549 OSS Projects crawled from Sourceforge portal. This covers about 84% of the total of 160,141 OSS Projects registered at the portal in the month of January There are more than 27 parameter being recorded which are project s name, audience, audience count, database environment, database environment count, developer count, development status, development status count, number of download, filename and file size, license, license count, operating system, operating system count, programming language, programming language count, review count, topic, topic count, translation, translation count, user interface, and user interface count. The result of this datamining process are 9 success rules that may be applied by initiators and contributors of OSS Project in order to increase the probability of success of their projects. The details of the guidelines is shown in Section 4. Future work of this research include expanding the experiment to cover other portal such as launchpad.net, Google code and Freshmeat. Other possible exploration is by using more advanced learning rule other than the association rule. Replication Case Study of Open Source Development, IEEE Transaction on Software Engineering Vol. 31 No. 6, June 2005, [5]. A.W.R. Emanuel, R.Wardoyo, J.E. Istiyanto, K. Mustofa, Success Factors of OSS Projects from Sourceforge using Datamining Association Rule, Proceeding of the 2 nd International Conference on Distributed Frameworks and Applications (DFmA), 2010, [6]. V.K. Gurbani, A. Garvert, J.D. Herbsleb, A Case Study of Open Source Tools and Practices in Commercial Setting, Proceeding of the fifth Workshop on Open Source Software Engineering, 2006, 1-6. [7]. P.L. Li, J. Herbsleb, M. Shaw, Finding Predictors of Field Defects for Open Source Software Systems in Commonly Available Data Sources: a Case Study of OpenBSD, Proceeding of 11th IEEE International Software Metrics Symposium, 2005, 32. [8]. G. von Krogh, S. Spaeth, S. Haefliger, Knowledge Reuse in Open Source Software: An Exploratory Study of 15 Open Source Projects, Proceeding of 38th Hawaii International Conference on System Sciences, 2005, 198b [9]. A. Mockus, R.T. Fielding, J. Herbsleb, Two Case Studies of Open Source Software Development: Apache and Mozilla, ACM Transaction on Software Engineering and Methodology Vol. II No. 3, Juli 2002, [10]. A. Mockus, R.T. Fielding, J. Herbsleb, A Case Study of Open Source Software Development: The Apache Server, ACM ICSE, 2000, [11]. E.S. Raymond, The Cathedral and the Bazaar. Knowledge, Technology & Policy, vol. 12 no. 3 pp , [12]. S. Spaeth, M. Stuermer, Sampling in Open Source Development: The Case for Using the Debian GNU/Linux Distribution, Proceedings of the 40th IEEE Hawaii International Conference on System Sciences, 2007, 166a. Acknowledgments The authors of this paper would like to thank Maranatha Christian University ( that provide the funding for this research and Department of Computer Science and Electronics at Gadjah Mada University ( which provide guideline and technical assistance for the research. References [1]. R.Agrawal, R. Srikant, Fast Algorithm for Mining Association Rule, Proceeding of 20th International Conference Very Large Database, 1994, pp [2]. A. Capiluppi, J. F. Ramil, Studying the Evolution of Open Source Systems at Different Levels of Granularity: Two Case Studies, Proceeding of the 7th International Workshop of Principles of Software Evolution, 2004, [3]. S. Christley, G. Madey, Analysis of Activity in the Open Source Software Development Community, Proceeding of the 40th IEEE Annual Hawaii International Conference on System Sciences, 2007, 166b. [4]. T.T. Dinh-Trong, J.M. Bieman, The FreeBSD Project: A

102 81 Scalable Contents Delivery System with Dynamic Server Deployment Yuko KAMIYA 1, Toshihiko SHIMOKAWA 2, Fuminori TANIZAKI 3 and Norihiko YOSHIDA 4 1 Graduate School of Information Science, Kyushu Sangyo University Matsukadai Higashi-ku, Fukuoka , Japan 2 Faculty of Information Science, Kyushu Sangyo University Matsukadai Higashi-ku, Fukuoka , Japan 3 NTT West Fukuoka Blanch Hakataekihigashi Hakata-ku, Fukuoka , Japan 4 Information Technology Center, Saitama University 255 Shimoookubo Sakura-ku, Saitama , Japan Abstract On providing broadband contents, to provide enough network bandwidth is an important. Existing Contents Delivery Network has mainly focused on increasing network bandwidth statically. Therefore, it is not flexible. In this paper, we propose Soarin, a novel contents delivery system to increase network bandwidth dynamically by deploying delivery servers in a wide area. Moreover Soarin can use various server deployment policy to deploy delivery servers, it can decide which server is suitable for content distribution. We call the criterion server deployment policy. We also propose several kinds of server deployment policies for typical contents delivery services. Keywords: CDN, Virtual Machine, Server Selection 1. Introduction With the rapid spread of the Internet, we can use broadband network even at home. Content holders provide several kinds of broadband contents for all over the Internet. They require network bandwidth to provide these contents. CDN (Contents Delivery Network)[1] is used to large-scale contents delivery. [1] describes that a CDN has some combination of a content-delivery infrastructure, a request-routing infrastructure, a distribution infrastructure, and an accounting infrastructure. The content-delivery infrastructure consists of a set of "surrogate" servers that deliver copies of content to sets of users. In this paper, we call the surrogate server delivery server. CDN can increase network bandwidth so that delivery servers are distributed in a wide area all over the Internet. Before using CDN to deliver contents, contents provider estimate the amount of the access to provision the enough processing power and network bandwidth. However CDN cannot provide their services during overload. This is because CDN cannot increase its network bandwidth and processing power flexibly. Cloud computing makes it possible to increase processing power dynamically by increasing servers. However, current cloud-computing systems cannot increase network bandwidth. This is because it increases servers only in a local area. Servers have to be deployed in a wide area to increase network bandwidth. There are three problems in deploying servers in a wide area. These problems is composed of three main parts:(1) how to deploy the servers, (2) where to deploy the servers, (3) when to deploy the servers. There are some research contributions for wide area live migration [2][3]. The results of these researches can be used in server deployment in a wide area. However, the purpose of these studies focuses how to deploy the servers. We tackle the problem when to deploy the servers. In this research, we propose Soarin, a novel contents delivery system. Soarin can increase network bandwidth dynamically by deploying delivery servers in a wide area. Therefore Soarin is scalable. Moreover, Soarin can use various server deployment policies to deploy delivery servers. Therefore Soarin is flexible. As we mentioned above, there are still two problems about server deployment, where and when. Server deployment policy of Soarin is programmable, therefore Soarin can decide both of them using server deployment policy. Soarin

103 82 selects a suitable physical machine and executes new delivery server inside a virtual machine on the physical machine. Generally speaking, a criterion of selecting physical machine is different by contents holder. We call the criterion server deployment policy. We also propose several kinds of server deployment policies for typical contents delivery services. Our contributions consist of the following: Propose an architecture of Soarin, a flexible and scalable contents delivery system. Propose some server deployment policy, which are suitable for typical contents delivery. The remainder of the paper is organized as follows. Section 2 presents a related works, including an overview and problem of CDN. The proposed scalable contents delivery system: Soarin is introduced in Section 3 and Section 4 presents an evaluation of our system. Finally we conclude this paper in Section Related Works Content Delivery Network (CDN) has been proposed for large-scale contents delivery. 2.1 CDN CDN is widely used to large-scale contents delivery. CDN has some combination of a content-delivery infrastructure, a request-routing infrastructure, a distribution infrastructure, and an accounting infrastructure. The content-delivery infrastructure consists of a set of "surrogate" servers that deliver copies of content to sets of users. In this paper, we call the surrogate server delivery server. The delivery servers are deployed all over the Internet. These servers cache the contents from origin server using content-delivery infrastructure. The origin server stores the original contents. Requests from clients are redirected to their suitable deliver server by requestrouting infrastructure. Finally, users retrieve contents from the delivery server by content-delivery infrastructure. Examples of commercial CDNs are Akamai [4] and Limelight [5]. 3. Flexible Contents Delivery System with Dynamic Server Deployment As mentioned 1.2., existing CDN lacks flexibility to increase network bandwidth. In this research, we propose Flexible Contents Delivery System with Dynamic Server Deployment: Soarin. We use Server Proliferation [6][7] as a basis of Soarin. Server Proliferation is introduced in presents Soarin in detail. 3.1 Server Proliferation Server Proliferation realizes increasing and decreasing processing power and network bandwidth of server system dynamically. To realize this, it increases and decreases servers in a wide area dynamically. Figure 1. shows architecture of Server Proliferation. We introduce two types of the servers in Server Proliferation. One is Execution Server (ES) and the other is Distribution Server (DS). Server Proliferation deploys physical machines that are installed virtual machine monitor all over the Internet in advance. These physical machines execute virtual machines on them. We call these physical machines Execution Server. The other server is Deployment Server. DS stores HDD images of virtual machines. In Server Proliferation, services (ex. Web server, Streaming server and so on) are executed inside virtual machines. When a new virtual machine is required, a HDD image of virtual machine is distributed from DS to one of the ES. The distributed virtual machine is executed on the ES. In this architecture, DS can become a bottleneck. However, it is easy to use multiple DS. Therefore, it does not become a bottleneck. 2.2 Problems about Existing CDN Existing CDNs cannot increase its network bandwidth dynamically. It is because most of them cannot increase delivery server dynamically. If content holders want to deploy new servers, they have a their own demand to select new location of servers. However, existing CDN provides only built-in rules that may not meet content providers demand. Therefore we can say that existing CDNs lack flexibility to deploy distribution servers. Fig. 1 Architecture of Server Proliferation

104 83 Cloud computing makes it possible to increase servers dynamically. By increasing servers, it is possible to use the CPU and the network of the increased server. Thus processing power and network bandwidth is increased. Therefore cloud-computing systems can increase the processing power and network bandwidth. However, compared with processing power, it is difficult to increase network bandwidth. This is due to network bottleneck. Typical cloud-computing system is constructed in an idc (Internet Data Center). The uplink network of the idc may become network bottleneck of the cloudcomputing system. By contrast, Server Proliferation can increase both of processing power and network bandwidth. It is because it can deploy servers in a wide area; therefore, deployed servers can use different uplink network each other. As it turned out, it is possible to increase network bandwidth of the system. Server Proliferation uses virtual machine as a basis. It is because virtual machine is easy to increase and decrease dynamically. Moreover using virtual machine can reduce cost since physical machines can be shared with other system that uses virtual machines. It is possible to execute another virtual machine besides virtual machine executed by Server Proliferation. We can say that Server Proliferation is high-cost performance. 3.2 Soarin In this section, we describe Flexible Contents Delivery System with Dynamic Server Deployment: Soarin. Soarin realize increasing network bandwidth flexibly. Soarin increases distribution servers to increase network bandwidth. Soarin can increase distribution server anytime. It is possible to add network bandwidth even after content distribution is started. In addition, Soarin can decide when and where to increase distribution server flexibly. Upon increasing distribution servers dynamically, the new problem when and where to increase surrogates happens. In Soarin, distribution servers are constructed inside virtual machines, this problem is equal to a problem what criteria a physical machine to execute a virtual machine is chosen. We call this selection criterion Server Deployment Policy. We are studying about request navigation technology of CDN [8], and we clarify that there is a variety of selection criterion of request navigation. Hence it is obvious that there is a variety of server deployment policy, too. That is to say, server deployment policy is different by contents holder. Soarin can use various server deployment policies for contents holders requirements. The examples of the server deployment policy are discussed in detail in section 3.3. Soarin uses Server Proliferation to deploy distribution servers. Server Proliferation provides how to deploy a virtual machine dynamically. However it lacks capability to decide timing and location of deploying virtual machine. Therefore, we introduce Observation Server (OS) and Control Server (CS) in addition to Execution Server (ES) and Deployment Server (DS) in Server Proliferation. OS collects several kinds of metrics using server deployment policy. For example, it collects the CPU load and network traffic of ESs, calculates the distance from ESs to DSs. CS controls all over Soarin s system. CS selects ES to deploy new virtual machine using information from OS based on server deployment policy. Afterward CS directs DS to transfer HDD image of the virtual machine to the selected ES to deploy new distribution server. After that, CS directs ES to execute the virtual machine for new distribution server. Finally CS updates request navigation policy to use the new distribution server. Soarin uses Tenbin[8] as a request navigation system. 3.3 Typical Server Deployment Policies of Soarin As mentioned above Soarin is able to use various server deployment policies. Server deployment policy may differ by each content holder. In this section, we show five typical server deployment policies. It is also possible to use other policies and/or combine these five server deployment policies. 1.Distance between Execution Server and Deployment Server This policy uses the distance between Deployment Server and Execution Server as a criterion. Some metrics can be used to calculate the distance, number of hops, round trip time, and AS (Autonomous System) path length between DS and ES. DS measure these information periodically. OS collects these information from DS. This policy chooses the nearest ES from DS; therefore it may deploy new distribution server in a short time. As a result, it is possible to correspond to a sudden surge in the volume of request. 2.Processing Power of Execution Server This policy uses processing capacity of Execution Servers as a criterion. OS collects processing power information of each ESs in advance. OS collects load averages of ESs periodically. OS calculates processing capacity of ESs with the load average and ESs own processing power. Then OS chooses most powerful server. This rule can solve the lack of the processing power of distribution servers. 3.Network Bandwidth

105 84 This policy uses capacity of network bandwidth of ES as a criterion. OS collects network bandwidth of uplink of ESs in advance. OS collects usage of network bandwidth of uplink of ESs. OS calculates capacity of network bandwidth with these information. OS chooses the ES that has the broadest network bandwidth. Thus we solve a lack of network bandwidth. 4. Region where largest number of clients This policy uses number of clients per region as a criterion. Soarin can deploy new distribution servers, however if the server is far from clients, the increased network bandwidth cannot use effectively. Therefore this policy tries to select nearest ES for most of clients. This policy has to define boundary of region. Some definition is available. Examples are Country, continent, and Autonomous System (AS)[9]. AS is a unit of internetworking routing on the Internet, typically an Internet Service Provider or a very large organization. We can use IP address of clients to distinguish region. To distinguish the country from IP address, we can use information from RIR (Regional Internet Registry). A RIR is an organization that manages the allocation and registration of IP addresses with in a particular region of the world. There are five RIRs: ARIN, RIPE NCC, APNIC, LACNIC, and AfriNIC. ARIN manages North America, RIPE NCC manages Europe, the Middle East, and Asia, APNIC manages Asia and the Pacific area, LACNIC manages Latin America and the Caribbean Area, and AfriNIC manages Africa. We can recognize country of IP address from the RIR s information, and also recoginize that RIR assigns the IP address. If IP address of clients and ES are assigned from same RIR, it may exists near area. Generally speaking, the network bandwidth of the near area is wider. For example, it is rational to choose the ES on Europe but not on North America and or Africa toward the access from Europe. 5.Contents holder s opinion This policy uses contents holder s opinion as a criterion. For example, if economical cost is not same between ESs, there is a possibility that not only the performance but also the cost is a better criterion. At that time the contents holder wants to choose ES whose costs are cheap. On the other hand, some content holder wants to do traffic engineering. Accordingly, the content holder selects ES by manual. 4. Conclusions There are two problems in deploying servers in a wide area. One is where to deploy the servers, and the other is when to deploy the servers. In this paper we tackled the problem where to deploy the servers. In this paper, we proposed Soarin, a novel contents delivery system. Soarin can increase network bandwidth dynamically by deploying delivery servers in a wide area. Moreover Soarin can use various server deployment policy to deploy delivery servers, it can decide which server is suitable for content deliverer. We call the criterion server deployment policy. We also propose several kinds of server deployment policies for typical contents delivery services. Acknowledgments This research was supported in part by MEXT in Japan under Grants-in-Aid for Scientific Research on Priority Area , and by JSPS in Japan under Grants-in- Aid for Scientific Research (B) References [1] Day, M., Cain, B., Tomlinson, G. and Rzewski, P.: A Model for Content Internetworking (CDI), RFC3466 (2003) [2] Takahiro Hirofuchi, Hirotaka Ogawa, Hidetomo Nakada, Satoshi Itoh, Satoshi Sekiguchi, A Storage Access Mechanism for Wide-Area Live Migration of Virtual Machines, IPSJ SIG Technical Reports. [High Performance Computing] 2008(74), 19-24, (July, 2008) [3] Robert Bradford, Evangelos Kotsovinos, Anja Feldmann, and Harald Schiberg, "Live wide-area migration of virtual machines including local persistent state. ", Proceedings of the 3rd international conference on Virtual execution environments, pp (2007) [4] Akamai Technologies Inc: Akamai; The Business Internet. [5] Limelight: [6] Yuko Kamiya, Toshihiko Shimokawa, "Scalable Server Construction Method Based on Virtual Machine Transfer and Duplication," Proceedings of International Multi-Conference on Engineering and Technological Innovation IMETI 2008, Vol.2, pp (June, 2008) [7] Yuko Kamiya, Toshihiko Shimokawa, Norihiko Yoshida, "Scalable Server System Based on Virtual Machine Duplication in Wide Area", Proceedings of The 3rd International Conference on Ubiquitous Information Management and Communication, pp (January, 2009) [8] Toshihiko Shimokawa, Norihiko Yoshida, Kazuo Ushijima, Server Selection Mechanism with Pluggable Selection Policies, Electronics and Communications in Japan, Part (Fundamental Electronic Science), Vol.89, No.8, pp.53-61(july, 2006) [9] Rechter, Y., Li, T., A border gateway protocol 4(BGP-4), RFC1771(1995) Yuko Kamiya received the B. Information Science and the M. Information Science from Kyushu Sangyo University in 2006 and 2008 respectively, where she is currently working toward the Ph.D degree. Her research interests include load balancing and wide area networking. She is a member of Information Processing Society of Japan.

106 85 Toshihiko Shimokawa received the B.E. and M.E. degree in computer science and communication engineering and the Ph.D. degree in computer science from Kyushu University in 1990, 1992 and He is a professor of information science in the Department of Information Science at Kyushu Sangyo University since His research interests include parallel and distributed computation, and wide area networking. He is a member of Information Processing Society of Japan and The Institute of Electronics, Information and Communication Engineers. Fuminori Tanizaki joined NTT West in Since he joined NTT West, he has been engaged in Network & Server Operation. His research interests include IPv6, Video Streaming, DNS, and BSD UNIX. Norihiko Yoshida received the M. Eng. degree in mathematical engineering and information physics in 1981 from the University of Tokyo, and the Dr. Eng. degree in computer science and communication engineering in 1990 from Kyushu University. He is Professor of computer systems at the Department of Information and Computer Sciences since 2002, and Director of Information Technology Center since 2008, both of Saitama University. His research interests include system design methodologies, and parallel and distributed computation. He has been served on program committees and organizing committees of various international conferences and symposia. He is a member of IPSJ, IEICE, JSSST, ACM and IEEE.

107 A Survey on Performance Evaluation of Object Detection Techniques in Digital Image Processing 86 Mrs J. Komala Lakshmi 1, and Dr.M.Punithavalli 2 1 Assistant Professor, SNR SONS COLLEGE, Coimbatore , Tamil Nadu, India 2 Professor and Head, Department of Computer Science Dr.SNS Rajalakshmi College of Arts and sciences, Coimbatore, Tamil Nadu, India. Abstract In digital image processing, the performance evaluation means the analysis of parameters that improves the execution of the proposed system there by producing the optimized result. The image is defined as a Scene consists of objects of interest. To understand the contents of the image, one should know the objects that are located in the image. The shape of the object is a binary image representing the extent of the object. In Digital Image processing the shapes are represented and described in various methods.shape representation method results in a non numeric representation of the original shape (e.g.) a graph. So that the important characteristics of the shape are preserved. The shape description refers to the methods that result in a numeric descriptor of the shape and is a step subsequent to shape representation.skeletons are one such shape descriptors. The skeleton of a two-dimensional object is a transformation of the shape object into a one dimensional line introducing skeleton shape descriptors. Many operations like shape representation and deformation can be performed more efficiently on the skeleton than on the full object, as skeleton is simpler than the original object. The parameters such as thresholds, bounds and weights have to be tuned for the successful performance of the object recognition system. This paper provides an overview of estimating the parameters for performance evaluation of the object detection techniques, and a survey of Performance evaluation of junction detection schemes in digital image processing. Keywords: Performance analysis, Roc analysis, Performance Criteria, Parameter selection, Junction detection. 1. Introduction In any system that are newly developed, it is highly recommended to go for testing or sample execution. The output of the system with the user input data is compared or analyzed against the expected output with the system defined data. After such analysis,it is identified that some of the factors used in the system may affect or change the expected output.such factors need to be changed for the improved output or the result.those factors are called as parameters. Parameters are those combinations of the properties which suffice to determine the response of the system. Properties can have all sorts of dimensions, depending upon the system being considered; parameters are dimensionless, or have the dimension of time or its reciprocal [1]. This paper provides a summery of object detection techniques, the parameters involved,performance criteria investigated and evaluation of different junction detection schemes. 2. Background The performance analysis, more commonly today known as testing, is the investigation of a program's behavior using information gathered as the program executes (i.e. it is a form of dynamic program analysis, as opposed to static code analysis). The usual goal of performance analysis is to determine which sections of a program to optimize - usually either to increase its speed or decrease its memory requirement (or sometimes both). 2.1 Algorithmic efficiency. In computer science, efficiency is used to describe properties of an algorithm relating to how much of

108 various types of resources it consumes. The two most frequently encountered are 1. Speed or running time - the time it takes for an algorithm to complete, and 2. Space - the memory or 'non-volatile storage' used by the algorithm during its operation, but also might apply to 3. Transmission size or external memory such as required bandwidth or disk space. The process of making code as efficient as possible is known as Optimization and in the case of automatic optimization (i.e. compiler optimization) - performed by compilers (on request or by default) - usually focus on space at the cost of speed, or vice versa. To analyze an algorithm is to determine the amount of resources (such as time and storage) necessary to execute it. Most algorithms are designed to work with inputs of arbitrary length. Usually the efficiency or complexity of an algorithm is stated as a function relating the input length to the number of steps (time complexity) or storage locations (space complexity).algorithm analysis is an important part of a broader computational complexity theory, which provides theoretical estimates for the resources needed by any algorithm which solves a given computational problem. These estimates provide an insight into reasonable directions of search of efficient algorithms. Exact measures of efficiency are useful to the people who actually implement and use algorithms, because they are more precise and thus enable them to know how much time they can expect to spend in execution. 2.2Optimization techniques The fig 1 figurative representation depicts the various techniques available. 87 Finally a matching or testing is performed with that of the model and the proposed system is done. If the Proposed system produces the expected output/result then the selected parameters are the optimizing parameters and are said to support the performance of the proposed system. Otherwise the selected parameters are to be modified and new set of parameters have to be estimated for the better performance of the proposed system. Performances of various cornerness measures are discussed with respect to four performances of robustness: detection, localization, stability and complexity.[2]object recognition systems almost inevitably involve parameters such as thresholds, bounds and weights[3]. The selection of parameters is a critical one for the system to perform successfully. The manual method performs parameter estimation in an ad hoc way by trial and error. A combination of parameters is selected to optimize the objective function and the optimum is compared with the desirable result in the designer s perception and the selection is adjusted. This process is repeated until a satisfactory choice, which makes the optimum consistent with the desirable result, is found.[4] 3.1Learning model We describe how to model the appearance of an object using multiple views, learn such a model from training images, and recognize objects with it in fig 2. The model uses probability distributions to characterize the significance, position, and intrinsic measurements of various discrete features of appearance; it also describes topological relations among features. The features and their distributions are learned from training images depicting the modeled object.a matching procedure, combining qualities of both alignment and graph sub isomorphism methods uses feature uncertainty information recorded by the model to guide the search for a match between model and image [5]. Fig.1 3. Methodology For performing the performance analysis, first a model should be developed. Secondly, the proposed system with the selected parameters is designed.

109 Fig.2Learning a multiple view models from training images requires a clustering of the training images and a generalization of each cluster s content Learning appearance models Learning a multiple-view model from real images requires some means of comparing and clustering appearances. Although several researchers have clustered images rendered from CAD models and thus avoided the feature correspondence problem, only a few have clustered real images. Among them,[6] measures the similarity of an image pair as the proportion of matching shape features, whereas[7] use a vector clustering algorithm with fixed-length vectors encoding global appearance. 88 such keys may well be consistent with a number of global hypotheses Parameter Selection A fundamental component of the approach is the use of distinctive local features we call keys. A key[9] is any robustly extractable part or feature that has sufficient information content to specify a configuration of an associated object plus enough additional, pose-insensitive (sometimes called semi-invariant) parameters to provide efficient indexing. The local context amplifies the power of the feature by providing a means of verification Identifying the Junction 3.1.2Parameter estimation for optimal object recognition They use coordinate system from geometric features,did match quality measure, estimate feature mach probability, estimating aligning transformation, and derive matching procedure for learning model. 3.2Designing the proposed system A common method is to generate a variety of input images by varying the image parameters and evaluate the performance of the algorithm as algorithm parameters vary. Operating curves that relate the probability of mis parameter setting. Such an analysis does not integrate the performance of the numerous operating curves. This process involves 1.Identifing the object.2.select the key features of that object 3. Identify the junctions(keys) Identifying the Object The basic idea is to represent the visual appearance of an object as a loosely structured combination of a number of local context regions keyed by distinctive key features, or fragments.[8] A local context region can be thought of as an image patch surrounding the key feature and containing a representation of other features that intersect the patch. Now under different conditions (e.g. lighting, background, changes in orientation etc.) the feature extraction process will find some of these distinctive keys, but in general not all of them. Also, even with local contextual verification, The recognition technique is based on the assumption that robustly extractable, semi-invariant key features, which are subsequently verified in local context, can be efficiently recovered from image data[10]. More specifically, the keys must posses the following characteristics. First, they must be complex enough not only to specify the configuration the object, but to have parameters left over that can be used for indexing. Second, the keys must have a substantial probability of detection if the object containing them occupies the region of interest (robustness). Third, the index parameters must change relatively slowly as the object configuration changes (semi-invariance). From a computational standpoint, true invariance is desirable, and a lot of research has gone into looking for invariant features [11]. 4. Performance Analysis frame work Fig.3 The basic idea is to represent the visual appearance of an object as a l From the literature in a digital image, one can basically distinguish between two approaches as in fig 3.and each of them corresponds in some way to the processing of one component in the decomposition above. 1. The stochastic approach, which is based on the modeling of an image as a realization of a random

110 process. Usually, it is assumed that the image intensity derives from a Markov Random Field and, therefore, satisfies properties of locality and stationary, i.e. each pixel is only related to a small set of neighboring pixels and different regions of the image are perceived similar. This modeling is particularly adapted for texture images (thus to the processing or the component u2 in the previous decomposition) and has motivated numerous works on texture analysis and synthesis [12]. 2. The deterministic approach, whose main purpose is to recover the geometry of the image. 4.1Markov Random Field Approach A pioneering work on the recovery of plane image geometry is due to[13]. They did not directly address the problem of recovering missing parts in an image but rather tried to identify occluding and occluded objects in order to compute the image depth map. Their algorithm starts with the detection of the boundaries of image objects. The next step is the identification of occluded and occluding objects. To this aim, [13] had the luminous idea to mimic a natural ability of human vision to complete partially occluded objects, the so-called a modal completion process described and studied by the Gestalt school of psychology and particularly[14]. Fig 4.T-junction The theory is applied to a specific model of MRF recognition presented in[15].the process is based on supervised learning process. Correctness, instability and optimality are proposed as the three level criteria for evaluating the parameter estimates Skeleton based approach. Superiority of skeleton is that it contains useful information of the shape, especially topological and structural information. To have skeleton of a shape, first boundary or edge of the shape is extracted using edge detection algorithms [16] and then its skeleton is generated by skeleton extraction methods [17] Medial axis is a type of skeleton that is defined as the locus of centers of maximal disks 89 that fit within the shape [18]. We use medial axis as the skeleton of shape. In [19] they present an algorithm for automatically estimating a subject s skeletal structure from optical motion capture data without using any a priori skeletal model. In [20].Other researchers have worked on skeleton fitting techniques for use with optical motion capture data. In [21] describe a partially automatic method for inferring skeletons from motion. They solve for joint positions by finding the center of rotation in the inboard frame for markers on the outboard segment of each joint. The method of [22], like ours, works with distance constraints although they still rely on rotation estimates. They assume that the skeletal topology is known beforehand and use heuristics to test multiple possible marker assignments. Similar problems have also been studied in the biomechanics and robotics literatures. A few specific examples of methods for inferring information about a human subject s skeletal anatomy from the motions of bone or skin mounted markers can be found in [23]. In [24]they have published a survey of calibration by parameter estimation for robotic devices Parameter used. They determine the skeleton s topology and the locations of the connecting joints. Both are determined by minimizing the same quantity, called the joint cost. A joint between two segments in an articulated skeleton should maintain a constant distance from the markers in marker groups for both segments. To avoid excessive computational costs we only solve the all pairs joint optimization approximately, then once we know the skeleton topology we solve for just those joints more accurately. The parameter used is the junction cost and the performance criteria are Topological connectivity, Qualitatively accurate structure, and non-linear optimization. In[25]They consider extracted skeleton as a connectivity graph such that junctions are considered as graph nodes and skeletal curve segments is considered as graph edges1. Connectivity graph perfectly represents topology of the skeleton and structure of the shape. Fig. 5 shows a sample shape, its skeleton and its connectivity graph Topological Information We represent skeleton of a shape as a graph such that junction points are graph nodes and skeletal curve segments are graph edges. We call this graph

111 as connectivity graph of the skeleton. Therefore, we may have for any given shape, its skeleton and its connectivity graph. 90 sensitivity of performance to relevant factors such as the context of the edge. 4.4.Junction based approach Fig.5.Shape and connectivity graph Geometric Information In order to include geometric information into our shape representation method, we need to a feature which captures convexities, concavities, thickness and thinness of different parts of a shape. This is achievable using "radius function". This function is defined and can be computed for all skeletal points. Radius function R(p) for point p on the skeleton is the radius of the maximal inscribed disc touches the boundary of the shape [26]. In fact, variations of this function along the skeletal points create convex or concave parts of a shape. This is shown in Fig. 6. As seen in this figure, fixed values of radii from A to B create a flat part and increases of radii values from B to C create a convex part in the shape. A strength of the methodology is that it can be applied to any detection problem. The line detection example developed in this paper was for demonstrating the application of this methodology.this methodology has been partially adapted for performance evaluation of object recognition algorithms[29] and machine inspection algorithms[30].the key steps to applying this methodology to any algorithm are(i)converting the algorithm into a detection algorithm, (ii) choosing the appropriate signal variable to use as the threshold. Another appropriate example where our methodology could be used is the detection of corners and junctions[31] Evaluation of Junction Detection schemes In general,the frame work is described as follows in the fig 7. Fig.6.Geometric information 4.3Edge based approach. In [27],They design and define the Performance by asking the following questions. 1.How exactly does one define Performance?.Issues that are need to be addressed are 2.What image population is relevant? 3.Is the performance evaluated independent of the algorithm? 4.How are difference in performance measured? with previous work on quantitative performance evaluation is in edge detection and thresholding [28].Most of the papers present an analysis that is specific to edge detection.furthermore,the performance is finally a number,(e.g.) percentage of edge points detected,etc. There is little further analysis of the Fig 7 Junction detection schemes A number of different methods have been proposed to evaluate the various approaches to corner detection. The different methods can be classified into methods based on visual inspection, localization accuracy, and theoretical analysis [32]. Localization accuracy is another evaluation method and can be measured based on the correct projection of 3D scene points to 2D image points[33].since this method requires the precise knowledge of 3D points, the evaluation is restricted to simple scenes of, for example, polyhedral objects. The performance of various corner detectors can also be assessed by theoretical analysis[34]. Analytical studies are limited to particular configurations such as L-corners. Here we have introduced the method of ROC analysis in the context of junction detection. ROC analysis allows assessment of the capabilities of the

112 detectors over the full range of possible thresholds for every test image. Consequently, ROC-based evaluation results are not flawed by choice of a particular threshold, which can strongly bias the obtained results Biologically motivated scheme for robust Junction detection. In [35],the performance criteria is classified into two approches. 1.Based on threshold 2.Threshold free approach, where Receiver operator characteristic (ROC) analysis is used for a threshold-free evaluation of the different approaches ROC analysis. Recently[36] studied the performance of human observers for the detection of junctions in natural images. Comparing [37] Unlike local approaches as proposed in computer vision [38], the new scheme is based on a more global, recurrent long-range interaction for the coherent computation of contour responses. In[39] various models namely, Models of Recurrent Long-Range Interaction[40] A comprehensive overview of these different approaches can be found in [41] Decoding Population Codes. [42]Circular Variance Function. [43] Multi scale Processing for Junction Detection [44] have been discussed Applying ROC for the Evaluation of Different Junction Detectors ROC analysis allows characterizing different detectors over the full range of possible biases or thresholds. In virtually all junction detection schemes, some kind of thresholding is involved, and the detection performance crucially depends on the determination of the optimal threshold value. A threshold-free evaluation of different detectors as provided by ROC analysis allows separating the sensitivity of the detector from its threshold selection strategy. ROC analysis in general is based on ground-truth verification, that is, the comparison of a detection result with ground truth. Thus, the first step to apply ROC analysis for junction detection is the specification of ground truth junction points for each test image. For synthetic images, the ground truth position of junction points is known from the definition of the image or can be rather easily inferred from the gray-level variations. The ROC curve characterizing the detection 91 performance of the particular method is obtained by plotting the true-positive rates against the false positive rates.to sum up, ROC analysis of the performance of junction detection schemes involves the following five steps: 1. Selection of an input image and determination of the ground truth position of junction points 2. Application of a particular junction detection scheme to the image 3. Normalization of the junction responses to the range [0; 1] 4. Variation of a threshold in N steps from 1 to 0 and computation of the respective true-positive tp and false-positive fp rate 5. Plot of the ROC curve, that is, plotting tp against fp The free parameters of the approach are the number of thresholds N and the error radius rerr Comparison of junction detection schemes a)evaluation of Junction Detection Based on Feedforward vs.recurrent Long-Range Processing In order to focus on the relative merits of the recurrent long-range interactions for the task of corner as in fig 4,and junction detection, the proposed scheme is evaluated using two different kinds of input, namely the activity W θ of the longrange stage and the purely feed forward activity W 0 of the complex cell stage. Localization of Generic Junction Configurations. From the outset of corner and junction detection in computer vision, the variety of junction types has been partitioned into distinct classes like T-, L-, and W-junctions,[45], and more recently, Ψ-junctions [46]. In the first simulation we compare the localization accuracy of junction responses based on feed forward vs. recurrent long-range responses for L-, T-, Y-, W- and Ψ-junctions (Fig. 8). For all junction type, the localization is considerably better for the method based on the recurrent long-range interaction. Fig.8.Processing of Images.

113 They have also evaluated the junction detection performance on real world images, such as cubes within a laboratory environment (Fig. 9). At the complex cell stage, many false responses are detected due to noisy variations of the initial orientation measurement. These variations are reduced at the long-range stage by the recurrent interaction, such that only the positions of significant orientation variations remain. We have further employed ROC analysis for the thresholdfree evaluation of the detection performance. The results show a better performance of the recurrent approach over the full range of thresholds (Fig. 3), Fig.10.Top row.images ; Bottom row.rocanalysis 92 Fig.9. Processing of Images. b)evaluation of Detection Performance Compared to Other Junction Detection Schemes In this section the author compared the new scheme based on recurrent long-range interaction with two junction detection schemes proposed in computer vision that utilize only localized neighborhoods, namely the structure tensor [47] Both schemes compute the.first- or second-order derivatives of the image intensity values, respectively. For a fair comparison of methods one has to ensure that all junction detectors operate on (at least approximately) the same scale [48]The derivatives used in the two standard methods are therefore approximated by Gaussian derivatives whose standard deviations are parameterized to.fit the successive convolution of filter masks used to compute the complex cell responses. We show the results of the ROC analysis when applied to a number of artificial and natural images, particularly a series of cube images within a laboratory environment (Fig. 4), and a second set of images containing an artificial corner test image from [49] a laboratory scene from[50] and an image of a staircase (Fig. 5). For all images, the ROC curve for the new scheme based on recurrent long-range interactions is well above the ROC curves for the other schemes, indicating a higher accuracy of the new method. Fig. 11. Top row.images ; Bottom row.rocanalysis 5. Conclusions In digital image processing, Junctions plays a major role in various perceptual tasks, such as the determination of occlusion relationships for figureground separation, transparency perception, and object recognition, among others. This paper provides a overall summery of various junction detection techniques and its performance evaluation. The new approach showing the superior performance of both synthetic and camera images called Receiver operating characteristics is also discussed. This paper will be useful for all visualization users who wish to proceed further in the field of object detection and reconstruction in digital image processing. Acknowledgments The author would like to express her gratitude to Mr. Dr.V.Sengoden, Principal and secretary,snr SONS COLLEGE,COIMBATORE for providing necessary infrastructure and for his constant encouragement that led to improvise the presentation of the quality of the paper. References [1]." John D. Trimmer, 1950, Response of Physical Systems (New York: Wiley), p. 13.

114 [2],Zhiqiang Zheng, Han Wang *, Analysis of gray level corner detection Eam Khwang Teoh,School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore , Singapore Received 29 May 1998; received in revised form 9 October [3].Grimson,W.EL.(1990).Object Recognition by Computer-The Role of Geometric Constraints. MIT Press, Cambridge, MA. [4] STANZ.LI, Parameter Estimation for Optimal Object Recognition: Theory and Application_,School of Electrical and Electronic Engineering Nanyang Technological University, Singapore [5] ArthurR.,Pope DavidG.Lowe, David Sarnoff Learning Appearance Models for Object Recognition,, ResearchCenter,CN530,Princeton,J DeptofCo mputerscience,universityofbritishcolumbiavancouver, B.C, CanadaV6T1Z4. [6].PGros.Matchingandclustering:Twostepstowardsautom aticobjectmodelgenerationincomputervision.inproc.aaa IFallSymp.:MachineLearninginComputerVision,AAAIPress,1993. [7].M.Seibert,A.M Waxman.Adaptive 3-D object recognition from multiple views. IEEETrans.Patt.Anal.Mach.Intell.14: ,1992. [8]Andrea Selinger and Randal C. Nelson, ``A Perceptual Grouping Hierarchy for Appearance-Based 3D Object Recognition'', Computer Vision and Image Understanding, vol. 76, no. 1, October 1999, pp Abstract, gzipped postscript (preprint) [9]Randal C. Nelson and Andrea Selinger ``A Cubist Approach to Object Recognition'', International Conference on Computer Vision (ICCV98), Bombay, India, January 1998, Abstract, gzipped postscript, also in an extended version with more complete description of the algorithms, and additional experiments. [10]Randal C. Nelson, ``From Visual Homing to Object Recognition'', in Visual Navigation, Yiannis Aloimonos, Editor, Lawrence Earlbaum Inc, 1996, Abstract, [11]Randal C. Nelson, ``Memory-Based Recognition for 3-D Objects'', Proc. ARPA Image Understanding Workshop, Palm Springs CA, February 1996, Abstract, gzipped postscript [12] Ashikhmin, M. (2001), Synthesizing Natural Textures, Proc. ACM Symp. Interactive 3D Graphics, Research Triangle Park, USA, March 2001, pp [13] Nitzberg, M., Mumford, D. and Shiota, T. (1993), Filtering, Segmentation and Depth,Lecture Notes in Computer Science, Vol. 662, Springer-Verlag, Berlin. [14] Kanizsa, G. (1979), Organisation in Vision, New- York: Praeger. [15].Li,S.Z(1994).A Markov random field model for object matching under contextual constraints,in proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition,seattle,Washington [16] T. Bernier, J.-A. Landry, "A New Method for Representing and Matching Shapes of Natural Objects", Pattern Recognition 36, 2003, pp [17] H. Blum, A Transformation for extracting new descriptors of Shape, in: W. Whaten-Dunn(Ed.), MIT Press, Cambridge, pp [18] Serge Belongie, Jitendra Malik, Jan Puzicha, Shape Matching and Object Recognition using Shape Contexts,IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol.24, No. 24, pp , April [19] Kirk James F. O Brien,,David A. Forsyth, Skeletal Parameter Estimation from Optical Motion Capture Data,Adam University of California, Berkeley [20] J. Costeira and T. Kanade. A multibody factorization method for independently moving-objects. Int. J. Computer Vision, 29(3): , [21] M.-C. Silaghi, R. Plankers, R. Boulic, P. Fua, and D. Thalmann. Local and global skeleton fitting techniques for optical motion capture. In N. M. Thalmann and D. Thalmann, editors, Modelling and Motion Capture Techniques for Virtual Environments, volume 1537 of Lecture Notes in Artifi cial Intelligence, pages 26 40, [22] M. Ringer and J. Lasenby. A procedure for automatically estimating model parameters in optical motion capture. In BMVC02, page Motion and Tracking, [23] J. H. Challis. A procedure for determining rigid body transformation parameters. Journal of Biomechanics, 28(6): , [24] B. Karan and M. Vukobratovi c. Calibration and accuracy of manipulation robot models an overview. Mechanism and Machine Theory, 29(3): , [25] Hamidreza Zaboli, Shape Recognition by Clustering and Matching of Skeletons, Amirkabir University of Technology, Tehran, Iran zaboli@aut.ac.ir,mohammad Rahmati and Abdolreza Mirzaei Amirkabir University of Technology, Tehran, Iran {rahmati,amirzaei}@aut.ac.ir [26] Dengsheng Zhang, Guojun Lu, Review of shape representation and description techniques, Pattern Regocnition 37, 2004, pp [27] Tapas Kanungo, A methodology for quantitative performance evaluation of detection algorithms, Student member,ieee,m.y Jaisimha, Student member,ieee,john Palmer and robrt M.Haralick,Fellow,IEEE [28]E.SDeutschandJ.RFram,\Aquantitativestudyoftheorie ntationalbiasofsomeedgedetectorschemes",ieeetransacti onson,computers,volc-27,no.3pp ,march1978. [29]O.ICamps,L.GShapiro,andR.MHaralick,\Recognition usingpredictionandprobabilisticmatching",inproc.ofieee/ RSJ FifthInternationalConferenceonIntelligentRobotsRaleigh, NorthCarolina,July192,pp [30]B.RModayur,L.GShapiro,andR.MHaralick\Visualins pectionofmachinedparts"inproc.ofieeeconf.oncomputer Vision andpatternrecognitionchampaign,iljune192,pp [31]C.HTehandR.TChin, On the detection of dominant points on digital curves Intelligence,vol1pp ,1989 [32] Schmid, C., Mohr, R., & Bauckhage, C. (2000). Evaluation of interest pointdetectors. Int. J. Comput. Vision, 37(2), [33] Coelho, C., Heller, A., Mundy, J. L., Forsyth, D., & Zisserman, A. (1991). An experimental evaluation of

115 projective invariants. In Proc. DARPA- ESPRIT,Workshop on Applications of Invariants in Computer Vision, Reykjavik, Iceland,(pp ). [34] Deriche, R., & Giraudon, G. (1990). Accurate corner detection: An analytical study. In Proc. 3rd Int. Conf. Computer Vision (pp ). Los Alamitos, CA: IEEE Computer Society Press. [35],Thorsten Hansen and Heiko Neumann, A Biologically Motivated Scheme for Robust Junction Detection Univ. Ulm, Dept. of Neural Information Processing, D Ulm, Germany,(hansen,hneumann)@neuro.informatik.uniulm.de,They propose a biologically motivated approach to junction detection [36] McDermott, J. H. (2002). Psychophysics with junctions in real images. J. Vis.,2(7), 131a. [37] Zucker, S.W., Dobbins, A.,&Iverson, L. A. (1989). Two stages of curve detection suggest two styles of visual computation. Neural Comput., 1, [38] Harris, C. J. (1987). Determination of ego-motion from matched points. In Proc Third Alvey Vision Conference (pp ). Cambridge, UK. [39] Thorsten Hansen, Heiko Neumann Neural Mechanisms for the Robust Representation of Junctions,,Giessen University, Department of Psychology, D Giessen, Ulm University, Department of Neural Information Processing, D Ulm, Germany [40] Grossberg, S., & Mingolla, E. (1985a). Neural dynamics of form perception:boundary completion, illusory figures, and neon color spreading. Psychol. Rev., 92, [41] Neumann, H., & Mingolla, E. (2001). Computational neural models of spatial integration and perceptual grouping. In T. F. Shipley & P. J. Kellman (Eds.), From fragments to objects: Segmentation and grouping in vision (pp ). Amsterdam: Elsevier Science. [42] Oram, M.W., F oldi`ak, P., Perret, D. I.,&Sengpiel, F. (1998). The ideal homunculus :Decoding neural population signals. Trends Neurosci., 21(6), [43] Gilbert, C. D.,&Wiesel, T. N. (1989). Columnar specificity of intrinsic horizontal and corticocortical connections in cat visual cortex. J. Neurosci., 9(7), [44] Lindeberg, T. (1998). Feature detection with automatic scale selection. Int. J.Comput. Vision, 30(2), [45] Huffman, D. A. (1971). Impossible objects as nonsense sentences. In B. Meltzer&D. Michic (Eds.), Machine intelligence 6 (pp ). Edinburgh: Edinburgh University Press. [46] Adelson, E. H. (2000). Lightness perception and lightness illusions. In M. S. Gazzaniga (Ed.), The new cognitive neurosciences (2nd ed., pp ). Cambridge, MA: MIT Press. [47] F orstner, W. (1986) A feature based correspondence algorithm for image matching. In Int. Arch. Photogramm. Remote Sensing, volume 26, pp [48] Lindeberg, T. (1998) Feature detection with automatic scale selection. Int. J. Comput. Vision, 30(2): [49] Smith, S. and Brady, J. (1997) SUSAN a new approach to low level image processing. Int. J. Comput. Vision, 23(1): [50] Mokhtarian, F., & Suomela, R. (1998). Robust image corner detection through curvature scale space. IEEE Trans. Pattern Anal. Mach. Intell., 20(12), The author Dr.M.Punithavalli 2, has finished her B.Sc Computer science from Dr. G.R.D College of Science, during 1991.She has completed her M.Sc Computer Science from Avinashilingam Deemed University during 1994.She has been Awarded her M.Phil Computer Science,Bharathiar University. During 2000.She Has been Awarded the Doctoral Degree from Alagappa University, Karaikudi. during 2007.She has Sixteen years of teaching experience in Collegiate service.she is currently working as Director and Head, Department of Computer Science, Dr.SNS Rajalakshmi College of Arts and Sciences, Coimbatore. She has presented her papers in seven national conferences and Ten international conferences.she has presented papers in more than five international journals and national journals.she also contributed five individual book publications to the computer science students.she has organised five national level, Symposiums. 40 M.phil candidates have been awarded under her guidance. Currently 12 members are doing their M.Phil Degree.Ten research scholars have been pursuing their Doctoral degree under her guidance and five scholars have been awarded. In short she dedicates her carrier to the Computer science students and the Emerging computer world. The author Mrs J.Komala Lakshmi 1, has finished her B.Sc Mathematics from SeethaLakshmi Achi College for women,pallathur, during 1995.She has completed her M.C.A from J.J.college Of arts and Sciences, Pudhukottai during 1998.She has been Awarded her M.Phil Computer Science,Bharathiar University during 2008.She Has been Awarded with the Silver meadal for her academic Excellence in M.C.A from the Institution from. She has Nine years of teaching experience in Collegiate service. She is currently working as Assistant Professor, Department of Computer Science, SNR SONS COLLEGE, Coimbatore. She has presented her papers in three international conferences.she has published three papers in threee international journals.she also contributed a book publications to the computer science students. She is a Graduate Student Member of IEEE. She has been awarded with the recognition from Who s Who in the World 2010, 28 th Edition for her notable contribution to the society.

116 95 Optimal Provisioning of Resource in a Cloud Service Yee Ming Chen 1 Shin-Ying Tsai Department of Industrial Engineering and Management, Yuan Ze University 135 Yuan-Tung Rd., Chung-Li, Tao-Yuan, Taiwan, ROC. Abstract Cloud service allows enterprise class and individual users to acquire computing resources from large scale data centers of service providers. This cloud service is more involved in purchasing and consuming manners between providers and users than others. However, Cloud service providers charge users for these services. Specifically, to access data from their globally distributed storage edge servers, providers charge users depending on the user s location and the amount of data transferred. User applications may incur large data retrieval and execution costs. Therefore, optimizing execution time, the cost arising from data transfers between resources as well as execution costs should be taken into account. In this paper, we present a discrete Particle Swarm Optimization (DPSO) approach for tasks allocation. We construct application Amazon EC2 as an example and simulation with Cloud based compute and transmission resources. Experimental studies illustrate that the proposed method is more efficient and surpasses those of mathematical programming and reflecting the actual benefit of saving with the total cost as well as tasks allocation. Keywords: Particle Swarm Optimization, Resource Allocation, Cloud service provider. 1. Introduction Cloud computing is a modality of computing characterized by on demand availability of resources in a dynamic and scalable fashion. The term resource here could be used to represent infrastructure, platforms, software, services, or storage. Cloud computing services allow users to lease computing resources from large scale data centers operated by service providers. Using cloud services, cloud users can deploy a wide variety of applications dynamically and on-demand. Most cloud service providers use machine virtualization to provide flexible and cost effective resource sharing. The cloud service provider is responsible to make the needed resources available on demand to the cloud users. It is the responsibility of the cloud service provider to manage its resources in an efficient way so that the cloud user needs can be met when needed at the desired Quality of Service (QoS) level[1]. Recently, many companies, such as Amazon, Google and Microsoft, have launched their cloud service businesses. Most cloud service providers use machine virtualization techniques to provide flexible and cost-effective resource sharing among users. Virtual machine(vm)instances normally share physical processors and I/O interfaces with other instances. It is expected that virtualization can impact the computation and communication performance of cloud services. Although most commercial providers present VM performance criteria to customers, it is difficult for management systems to assure VMs of their minimize execution cost or maximum assigned resources. If the tasks of VMs, for example, suddenly change from idle to active, the locations of VMs cannot be optimized again to meet the change[2]. In this paper, we propose meta-heuristic optimization approach based on Particle Swarm Optimization (PSO) for finding the near optimal tasks allocation with reasonable time. The approach is to dynamically generate an optimal task allocation so as to complete the tasks in a minimum period of time as well as utilizing the resources in an efficient way. The rest of the paper is organized as follows. Section 2 deals with some theoretical foundations related to tasks allocation model. In Section 3, we describe the proposed DPSO based algorithm in detail. Experimental results are presented in Section 4 and some conclusions and future works are provided towards the end. 2. Provisioning of Resources in a Cloud Environment Cloud computing services are often roughly classified into a hierarchy of as a service terms as following[3]: Infrastructure a s a Serv ice (IaaS) is providing general on-demand computing resources such as virtualized servers or various forms of storage (block, key/value, database, etc.) as metered resources. This can often be seen as a direct evolution of shared hosting with added ondemand scaling via resource virtualization and use-based billing.

117 96 Platform as a Serv ice (P aas) is providing an existent managed higher-level software infrastructure for building particular classes of applications and services. The platform includes the use of underlying computing resources, typically billed similar to IaaS products, although the infrastructure is abstracted away below the platform. Software as a Service (SaaS) is providing specific, already-created applications as fully or partially remote services. Sometimes it is in the form of web-based applications and other times it consists of standard nonremote applications with Internet-based storage or other network interactions. EC2 and other server clouds follow an IaaS model, in which the cloud users rent virtual servers and selects or controls the software for each virtual server tasks[4]. Every cloud service providers might have a unique way of managing and tasks allocation must ensure that they do not conflict with the resource owner's policies. In the worst-case situation, the cloud service providers might charge different prices to different cloud users for their resource usage and this might vary from time to time. Mathematical programming approaches [5] using column generation or branch-and-bound techniques can solve the tasks allocation problem[6]. However, the general n- processor tasks allocation has been found to be NPcomplete[7]. Therefore, finding exact optimum solutions to large-scaled tasks allocation problem is computationally prohibitive. The development of meta-heuristic optimization theory has been flourishing during the last decade. Particularly, with its sound exploration ability of both global and local optimal solutions, some new search techniques involving nature-inspired meta-heuristics have become the new focus of resource allocation research. As mentioned in [8] scheduling is NP-complete. Metaheuristic methods have been used to solve well-known NP-complete problems. Efficient Meta-heuristic methods, which are used frequently, are simulated annealing (SA) [9], genetic algorithm (GA) [10], ant colony optimization (ACO) [11] and particle swarm optimization (PSO)[12]. In this study, we consider the tasks allocation with the following scenarios(figure 1). The processors in the system are heterogeneous and they are capacitated with various units of memory and processing resources. Hence, a task will incur different execution cost if it is executed on different processors. On the other hand, all of the communication links are assumed to be identical and some communication cost between two tasks will be incurred if there is a communication need between them and they are executed on different processors. Figure 1 The framework of tasks allocation process In this paper, a version of discrete particle swarm optimization (DPSO) is proposed for cloud service provider s tasks allocation and the goal of allocation is to minimize the execution cost and communication cost mentioned above simultaneously. 2.1 Tasks allocation Model The Tasks allocation model [13,14]is an integer program with a quadratic objective function (1) which represents the total execution cost and communication cost, respectively. t p t 1 t p Min C( X) ec cc (1 x x ) i 1k 1 ik x ik i 1j i1 ij k 1 ik jk (1) Constraints: n k 1 t i1 t i1 x 1, 1,2, t (2) ik i, r x R, 1,2, p (3) i i ik m x ik k M k k,, 1,2, p (4) k, x ik (0,1) (5) Constraint (2) states that each task should be allocated to exactly one processor. Constraints (3) and (4) ensure that processing resource and the memory capacity of each processor is no less than the total amount of resource demands of all of its allocated tasks. The last constraint (5) guarantees that x ik are binary decision variables. As

118 97 mentioned in the previous section the goal of the tasks allocation is to minimize the total execution cost and communication cost simultaneously. 3. Proposed Discrete Particle Swarm Optimization Algorithm In this section we propose a version of discrete particle swarm optimization for tasks allocation. Particle needs to be designed to present a sequence of tasks in available cloud service providers. Also the velocity has to be redefined. Details are given what follows. In our method solutions are encoded in a t p matrix, called position matrix, in which p is the number of available processors at the time of allocation and t is the number of tasks. The position matrix of each particle has the two following properties: 1) All the elements of the matrices have either the value of 0 or 1. In other words, if X id is the position matrix of i-th particles in a d-dimensional space, then: X id t, p (0,1) 2) In each row of these matrices only one element is 1 and others are 0. In position matrix each row represents a task allocation and each column represents allocated tasks in a processor. VelocityV id of each particle is considered as a t p matrix whose elements are in range[ V max, V max]. Also Pbest and nbest are t p matrices and their elements are 0 or 1 as position matrices. pid represents the best position that i- th particle has visited since the first time step and p gd represents the best position that i-th particle and its neighbors have visited from the beginning of the algorithm. In this paper we used star neighborhood topology for p gd. In each time step p id and p gd should be updated: V new id X weight V new id old id 1 ( t, p) 0 C rand1 ( pid X id ) C2 rand 2 ( p X 1 gd id if V new id ( t, p) max{ V otherwise new id ( t, p)} In (6) id ( t, p) is the element in t-th row and p-th column of the i-th velocity matrix in the updated time step of the algorithm and X new id ( t, p) denotes the element in t- th row and p-th column of the i-th position matrix in the updated time step. C 1 and C 2 are positive acceleration V new (6) ) (7) constants which control the influence of P id and P gd on the search process. Also rand 1 and rand 2 are random values in range [0, 1] sampled from a uniform distribution. weight which is called inertia weight was introduced by Shi and Eberhart [7] as a mechanism to control the exploration and exploitation abilities of the swarm. Usually w starts with large values (e.g. 0.9) which decreases over time to smaller values so that in the last iteration it ends to a small value (e.g. 0.1). Eq. (7) means that in each row of position matrix value 1 is assigned to the element whose corresponding element in velocity matrix has the max value in its corresponding row. If in a row of velocity matrix there is more than one element with max value, then one of these elements is selected randomly and 1 assigned to its corresponding element in the position matrix. The pseudo code of the proposed DPSO algorithm is stated as follows: Create and initialize a t p -dimensional swarm with P particles repeat for each particle i=1,,p do if f ( X id ) f ( p id ) then // f( ) represent the fitness Pid X id ; function of Eq.(1) end if f ( Pid ) f ( Pgd ) then Pgd P id ; end end for each particle i=1,,p do update the velocity matrix using Eq. (6) update the position matrix using Eq. (7) end until stopping condition is true; 4. Experimental results In this section, we will present the experimental results and comparative the computational performance. The platform for conducting the experiments in a PC with Dual Core Processor GHz CPU and 1.75GB RAM. All programs are coded in Java programming language in Borland JBuilder We give a formal description of our tasks allocation model. We start with a description of a cloud infrastructure. Then, we formalize user tasks and allocation of tasks on the cloud infrastructure. In our

119 98 model, we represent a cloud as a connected graph of networked computation nodes. We assume that there exists a communication link between each pair of nodes. We also assume that each link has an individual bandwidth and the data transfer on one link does not affect the other links. A node n corresponds to a computing entity like a physical or a virtual machine. An edge e is a communication link between two nodes. Figure 2 shows an example of a cloud. The cloud is depicted by the directed acyclic graph (DAG). The nodes contain tasks by users submit to be executed on the cloud. The upper part of the node, ec, represent task execution cost. The numbers on the edges represent the communication cost of bandwidth links. ec 1k ec 4 k ec 6k Processors EC2 Standard Instance P1 P2 P3 P4 P5 P6 P7 P8 Table1. Amazon EC2 Standard Instance Memory(M) 1.7GB 1.7GB 15GB 7.5GB 7.5GB 1.7GB 1.7GB 1.7GB CPU(GB) 8.0GB~9.6G B 8.0GB~9.6G B 8.0GB~9.6G B 5.0GB~6.0G B 5.0GB~6.0G B 4.0GB~4.8G B 4.0GB~4.8G B 4.0GB~4.8G B Executed cost(ec) $0.12~$0.1 4 $0.12~$0.1 4 $0.96~$1.1 1 $0.48~$0.5 2 $0.48~$0.5 2 $0.12~$0.1 4 $0.12~$0.1 4 $0.12~$0.1 4 ec 2k ec 3k ec 5 k ec 7k Figure 2 the directed acyclic graph of tasks To simulate our proposed DPSO algorithm for interconnection tasks graph in figure 2, we have used the data set of Amazon EC2 Standard Instance are shown in Table 1. The stopping criterion in DPSO is the number of generations such that no improvement is obtained in the value of fitness function (figure 3).The achieve results of eight tasks allocation are shown in Table 2. ec 8k 4.1 Comparative performances In this section, we present the comparative performances between the proposed DPSO and mathematical programming(table 3). The parameter values used in both of DPSO and mathematical programming LINGO are optimally tuned by intensive preliminary experiments to let the competing algorithms perform at the best level. To be specific, the parameter setting used by DPSO is (number of particles=15, c1=1,c2=3) and cc ij Table 3. Comparison of the performance for various tasks allocation Quantity Heuristics Math. programming DPSO LINGO Processors fitness Time Min Cost Time (sec) (sec) Tasks Figure 3 The convergence of DPSO for eight tasks allocation. Table 2 The eight tasks allocation solutions through DPSO Optimal Task 1 Task 2 Task 3 Task 4 Task 5 Task 6 Task 7 Task 8 Allocation Processor 8 Processor 7 Processor 7 Processor 1 Processor 2 Processor 2 Processor 7 Processor 6 Cost Total execution cost and communication cost $1.011

120 99 5. Conclusions This paper presented a version of Discrete Particle Swarm Optimization (DPSO) algorithm for tasks allocation. We used the heuristic to minimize the total cost of application tasks excution on Cloud computing environments. The performance of the proposed algorithm was compared with the mathematical programming method through carrying out exhaustive simulation tests and different settings. Experimental results show that the advantage of the DPSO algorithm is its speed of convergence and the ability to obtain faster and feasible allocation. As future work, the authors of the paper plan to carry out extended simulation studies that consider not only CPU time and memory space share but also network bandwidth as resources. Acknowledgments This research work was sponsored by the National Science Council, R.O.C., under project number NSC E References [1] W. Chung, R. Chang, A new mechanism for resource monitoring in Grid computing, Future Generation Computer Systems,Vol. 25. No.1.,2009,pp [2] M. Armbrust, A. Fox, R. Griffith, A. D. Joseph, R. H. Katz, A. Konwinski, G. Lee, D. A. Patterson, A. Rabkin, I. Stoica, and M. Zaharia. Above the Clouds: A Berkeley View of Cloud Computing. Technical Report UCB/EECS , EECS Department, University of California, Berkeley, Feb [3] I. Foster, Y. Zhao,I. Raicu, S. Lu, S. Cloud computing and grid computing 360-degree compared, Grid Computing Environments Workshop,2008, pp [4] Amazon Elastic Compute Cloud, [5] A. Ernst, H. Hiang, M. Krishnamoorthy, Mathematical programming approaches for solving task allocation problems, Proc. of the 16th National Conf. Of Australian Society of Operations Research, [6] G.H. Chen, J.S. Yur, A branch-and-bound-with-derestimates algorithm for the task assignment problem with precedence constraint, Proc. of the 10th International Conf. on Distributed Computing Systems, 1990, pp [7] Zs. Németh, V. Sunderam, Characterizing grids: Attributes, definitions, and formalisms, Journal of Grid Computing,Vol. 1. No.1,2003,pp [8] A. Abraham, H. Liu, M. Zhao, Particle swarm scheduling for work-flow applications in distributed computing environments, in: Metaheuristics for Scheduling: Industrial and Manufacturing Applications, in: Studies in Computational Intelligence, Springer Verlag, Germany, 2008, pp [9] A. Abraham, R. Buyya, B. Nath, Nature's heuristics for scheduling jobs on computational Grids, in: Proceedings of the 8th International Conference on Advanced Computing and Communications, Tata McGraw-Hill, India, 2000, pp [10] Y. Gao, H.Q. Rong, J.Z. Huang, Adaptive Grid job scheduling with genetic algorithms, Future Generation Computer Systems,Vol. 21,No. 1,2005, pp [11] A. Abraham, R. Buyya and B. Nath, Nature s heuristics for scheduling jobs on computational grids, Proc. of the 8th IEEE International Conference on Advanced Computing and Communications, India, 2000,pp [12] H. Liu, A. Abraham, An hybrid fuzzy variable neighborhood particle swarm optimization algorithm for solving quadratic assignment problems, Journal of Universal Computer Science, Vol.13, No.7, 2007, pp [13] P.Y.Yin, S.S. Yu,P.P. Wang, and Y.T. Wang, A Hybrid Particle Swarm Optimization Algorithm for Optimal Task Assignment in Distributed System, Computer Standards & Interfaces, Vol. 28, 2006, pp [14] P. Ruth, X. Jiang, D. Xu, and S. Goasguen. Virtual distributed environments in a shared infrastructure. Computer, Vol. 38, No. 5, 2005,pp Yee Ming Chen is a professor in the Department of Industrial Engineering and Management at Yuan Ze University, where he carries out basic and applied research in agent-based computing. His current research interests include soft computing, supply chain management, and system diagnosis/prognosis. Shin-Ying Tsai was a graduated student in the Department of Industrial Engineering and Management at Yuan Ze University, where she was studying basic and applied research in Cloud computing and heuristic algorithms. She now works in Gold Circuit Electronics as a Design Engineering.

121 100 Presenting a New Routing Protocol for increasing Lifetime of Sensor Network Vahid Majid Nezhad 1 and Bager Zarei 2 1 Department of Computer Engineering, Islamic Azad University, Shabestar Branch Shabestar, East-Azarbaijan, Iran 2 Department of Computer Engineering, Islamic Azad University, Shabestar Branch Shabestar, East-Azarbaijan, Iran Abstract Sensor Networks are systems with restrict resources that are scattered on a large scale with numerous deployment density and are used in aggregating data in a intelligent style. In this paper the concept of "efficient energy consumption" in routing protocol of sensor network, which causes to increase the lifetime of sensor network, has been studied and a taxonomy of various techniques for efficient energy consumption is presented. Also a new routing protocol is presented in a way that proposal protocol uses some special techniques to consume energy efficiently. After that we have simulated proposal routing protocol and compared its function with recent protocols and we have observed that proposal routing protocol increase the lifetime of sensor networks. Keywords: Sensor Network, Routing Protocol, Efficient Energy Consumption, Lifetime of Sensor Network. 1. Introduction Sensor networks are used as a new tool in applications like military, hospital control, domestic, etc. A sensor network consists of a lot of sensor nodes that they spread in an expected area and the purpose of this architecture is to aggregate desired information from environment and send them to the base station by the use of intelligent styles. In these styles it is tried to consume the energy of sensor nodes efficiently in order to increase the lifetime of sensor network. Routing protocols in sensor networks from network structure point of view can be divided into two main categories: flat and hierarchical. In flat routing protocols the concept of leader node dose not exists and all nodes are same. SPIN [1] and GBR [2] are some examples of this category. In hierarchical routing protocols the act of clustering and classification of nodes are done and some nodes are considered as leaders. From this group of protocols we can name LEACH [3], TEEN [4], GAF [5] and max-min length energy constrained [6]. Indeed there are other categories of protocols like data centric, location based, energy aware. In a way that each routing protocol can belong to one or several of mentioned groups. In data centric protocols, interesting issues are scattered in order to sensual tasks assign to nodes. There are two ways to scatter interesting issues: Sinks scatter their favorite topics to nodes. Nodes send a notice to sinks for their available data and wait for a request from the sink which is interested to that data. Data centric routing protocols need an attribute-base addressing mechanism in a way that in this routing user can query the characteristics of a phenomenon instead of query a node. For example the query that "which areas temperature is more than 70 F?" can be used instead of "what is the temperature of a specific node?" Some examples of this category are SPIN [1], GBR [2] and TEEN [4]. In location based routing protocols, location information are used for sending data to desired areas instead of sending them to whole network. GAF [5] is an example of this category. In energy aware routing protocols, the main focus is on efficient consumption of sensor nodes' energy to increase the lifetime of network. An example of this category is max-min length energy constrained [6]. The structure of this paper is that in second section some introductions about efficient consumption of energy and its effect on increasing the lifetime of network are presented. In third section by using some techniques for efficient energy consumption, a new routing protocol is developed and is simulated along with max-min length energy constrained routing protocol [6] and the results of

122 101 the simulation is presented. And in the last section a summery of this paper and conclusion is presented. 2. Efficient Energy Consumption in Routing Protocols The main topic that is discussed in this paper is limitation in energy that has effect on the lifetime of sensor network. The lifetime of sensor network is the time takes until a node or some nodes discharge their energy. So considering special characteristics of sensor networks such as limitation in nodes' energy and etc, we should use special routing protocols to increase the lifetime of sensor networks in a way that they differ from other routing protocols such as ad-hoc network routing protocols. In technical term these routing protocols that consider the limitation of sensor nodes' energy are called energy aware routing protocols. Energy aware routing protocols may use different techniques for efficient energy consumption in a way that efficient energy consumption causes to increase the lifetime of sensor network. As it is displayed in Fig. 1 these techniques are clustered in two groups: reduction of total energy consumption and fair energy consumption. so the number of transmission data packets decrease effectively and therefore the total energy consumption will be reduced. This technique in protocols like LEACH [3] is used. Increasing the delay is the defect of this design because cluster heads have to buffer received data packets for a while in order to fusion them. In "negotiation" technique before sending actual data packets to a node first that node will be queried, with the use of Meta data, to understand whether it is interested in receiving that data or not. In case that node had not received that data packet before and has the conditions of receiving that data, it will announce, with the use of Meta data, that it is interested in receiving that data. Then the actual data packet will be sent to that node. So transferring redundant data packets has been avoided effectively and it causes to reduce the total energy consumption. This technique has been used in some protocols like SPIN [1]. In the technique of "Turning on/off", it is tried to turn off the transmitter radio circuits of some sensor nodes when there is no need to their cooperation. As it is displayed in Fig. 2 the major part of sensor nodes energy consumes in the phase of radio communication [7], so this technique causes to reduce the total energy consumption. This technique has been used in some protocols like GAF [5]. Fig. 2 The amount of energy consumption in different functions of nodes [7] Fig. 1 Taxonomy of efficient energy consumption 2.1 Reduction of total energy consumption For this purpose we should decrease the number of transmitted data packets efficiently without loosing information. In other words, we prevent from transferring repeated and useless data packets. To do this, it is possible to use four techniques: 1) Data fusion 2) Negotiation 3) Turning on/off 4) Soft and Hard thresholds. In "data fusion" technique, which is usually perform by some special nodes like cluster head, received data packets which are overlapped with other data packets and represented a similar sensed phenomenon fused together In "Soft and Hard thresholds" technique by spreading both soft and hard thresholds, it is tried to avoid transmitting data packets that are not in expected range or indicate similar data. These thresholds are related to attributes of sensed data. The hard threshold allows the sensor nodes to transmit data only when their sensed attribute be in an expected range. Consequently the number of transmission decrease sufficiently. Once a node senses a new data, which is valid considering hard thresholds, it transmits that data only when it's attribute in comparison with pervious data changes by an equal amount or greater than the soft threshold. Subsequently soft threshold decreases the number of transmission if there would be a few or no changes in the attributes of data that is sensed. So this technique causes to reduce the number of transmitting data packets, consequently total energy consumption is also

123 102 reduced. This technique has been used in some protocols like TEEN [4]. 2.2 Fair energy consumption For this purpose the energy of sensor nodes should be consumed fairly. In other words in routing, sensor nodes should be cooperated in a way that the energy of some nodes does not finish earlier than other nodes. To do this, the "max-min path" technique can be used. It should be taken into consideration that consuming the energy of sensor nodes unfairly, causes that some nodes loose their energy earlier than others consequently it causes to reduce the lifetime of sensor network. Fig. 3 displays a diagram of node's energy consumption at a moment for an instance of this kind of network. While Fig. 4 displays a diagram of fair energy consumption of sensor nodes [2]. As you see total energy consumption in both diagrams are the same but in figure 4 sensor nodes have consumed energy more fairly so the energy of some nodes do not finish fast consequently the life time of sensor network increases. In "max-min path" technique when there is a data packet for sending, that data packet would be sent from a path which the minimum energy of that path's nodes is more than the minimum energy of other candidate path's nodes. So this technique causes that nodes which have lower energy do not take part in routing consequently their energy will not finish soon. This technique has been used in some protocols like max-min length energy constrained [6]. Fig. 3 Nodes' energy while inefficient energy consumption [2] Fig. 4 Nodes' energy while efficient energy consumption [2] 3. Presenting a new routing protocol In this section we have applied the "Soft and Hard Thresholds" and the "Turning on/off" techniques, which are presented in section 2, on max-min length energy constrained protocol [6] and we have developed a new routing protocol and simulated it. Indeed you should take it into consideration that max-min length energy constrained protocol [6] is equipped with technique of "max-min path" in advance. In our proposal protocol first, in order to apply the "Soft and Hard Thresholds" technique sink sends two parameters (T s for soft threshold and T h for hard threshold) to sensor nodes. After receiving these parameters, sensor nodes only transmit the data packets which satisfy these thresholds. Second, in order to apply the Turning on/off" technique clustering is done. Clustering in our proposal protocol is done in this way that if we consider the radio transmission domain of nodes as L meter, we can create the clusters as a square with sides of L/2.82 which starts from the sink. It guarantees that a cluster head in a cluster is able to communicate with other cluster heads in neighbor clusters. Indeed a technique without any need to GPS is also presented in [8] for creating clusters with constant, equal and symmetrical shapes without overlapping. We used this technique too. After clustering, in each cluster the node which contains the maximum energy is chosen as the cluster head. After that each cluster head order its members to turn off and only itself remain on the cluster to cooperate in routing jobs. After a while that routing of data packets have been done, each cluster head order its members to turn on so that all nodes participate at the phase of choosing the new cluster heads. Considering that the majority of energy will be consumed in communication phase and since the majority of nodes will be turned off, consequently the total energy consumption will be reduced. Indeed a lot of turned off nodes May causes an interruption in sensing jobs. So regarding spreading density of nodes we can consider an R parameter in a way that each cluster head while ordering "turn off", choose R nodes to have chance remains "turn on". For simulation we have used Glomosim [9] simulator. For this purpose we have considered a network which includes 500 nodes that are spread in a 400*400 m2 area randomly and the domain of radio transmission of nodes is 40 m. Each sensor node has a randomly initial energy between 1000 mj to mj. In this simulation the nodes use the protocol in MAC layer. The size of packets is considered 64 byte and we produce 800 data packet randomly in nodes so that the packets contain random temperatures between 1 and 100. Indeed for simulation it is supposed that sensor network monitor temperature

124 103 changes, T h (hard threshold) is between 20 and 80, and T s (soft threshold) is 10. Eq. (1) and (2) show the amount of consumed energy while sending and receiving k bit data in d meter distance. Indeed E_elec is the necessary energy for executing circuits/orbits and switching action among bites that usually is 50 nj/bit and E_amp is the necessary energy for amplifies radio for radio transmitting, that usually is 100 pj/bit/m2. 2 E _ Tx( k, d) E _ elec* k E _ amp* k * d (1) E _ Rx( k, d) E _ elec * k (2) Next parts are presented for investigation of "the fairness factor", "the number of received packets by sink" and "the average of remaining energy" about our proposal protocol and max-min length energy constrained protocol [6]. 3.1 Fairness factor In simulation we have used f for fairness factor which is calculated through Eq. (3) and it is the agent of fair energy consumption among nodes and its domain is between 0 and 1. Its optimal value is 1 and as it gets closer to 1, the energy consumption would be fairer among nodes. n is the number of nodes that in our simulation we have considered it 500. n n 2 2 (3) f ( ( E i )) ( n ( E i ) ) i 1 i 1 In this part of simulation we have compared our proposal protocol with max-min length energy constrained protocol [6] from f factor point and we have shown the results in Fig. 5. As shown in Fig. 5 at the beginning of simulating, because the energy distribution is not the same among nodes, f factor is low in both of them. But in course of time with performing both protocols, f factor has increased. Also our proposal protocol performs a little bit better and the reason is that max-min length energy constrained protocol [6] only uses "max-min path" technique. And this technique by itself causes the increase of f factor. But our proposal protocol not only uses "max-min path" technique but also uses "Turning on/off" technique that in the latter after clustering the state of nodes which contain less energy changes to turn off mode. So their energy consumption is prevented. And it causes the increase of f factor. Indeed the more R parameter increases, f factor decreases slightly. And it inclines to max-min length energy constrained protocol [6]. 3.2 The number of received packets by sink In this part of simulation we have compared our proposal protocol with max-min length energy constrained protocol [6] from the point of the number of received packets by sink. Simulating result is demonstrated in Fig. 6. Indeed it should be considered that naturally in our proposal protocol the number of sensed data packets is more than received data packets by sink, since some of the sensed data packets are assigned to nodes which are turned off and no routing operation is done on them. Indeed as shown in Fig. 6 by increasing the R parameter, the rate of sensing tasks increases and vice-versa. Also some data packets are omitted by applying "Soft and Hard thresholds" technique, that this omission is effective because omitted data packets are related to unacceptable and repeated information. Fairness Factor (f ) Number of received packets by Sink Number of sent packets Fig. 5 The fairness factor Number of sent packets Max-Min Length- Energy-Constrained [6] Proposal Protocol with R=0 Proposal Protocol with R=1 Max-Min Length- Energy-Constrained [6] Proposal Protocol with R=0 Proposal Protocol with R=1 Fig. 6 The number of received packets by Sink 3.3 The average of remaining energy In this part of simulation we have compared our proposal protocol with max-min length-energy-constraint routing protocol [6] from average of remaining energy point and the results are shown in Fig. 7. As it can be seen in Fig. 7 the amount of total energy consumption in our proposal protocol has decreased drastically. One of the major reasons of this event is omitting unacceptable or repeated data packets. Another major reason is the state of turned off unnecessary nodes and preventing from consuming their energy.

125 104 Average of remaining energy (mj) Conclusions Number of sent packets Max-Min Length- Energy-Constrained [6] Proposal Protocol with R=0 Proposal Protocol with R=1 Fig. 7 The average of remaining energy of nodes In this paper, first we discussed energy limitation in sensor networks and it relation with lifetime of sensor network. Then presented two ways for increasing the lifetime of sensor network: 1) reduction of total energy consumption 2) fair energy consumption. Then we discussed different techniques of each mentioned ways. We also stated that a suitable routing protocol should use techniques of both ways at the same time. Also, by using some techniques, we expand max-min length-energy-constraint routing protocol [6] and we presented a new routing protocol that increased the lifetime of sensor network through reduction of total energy consumption and fair energy consumption. At the results of simulation demonstrated, this new routing protocol helped to efficient energy consumption in a way that efficient energy consumption causes to increase the lifetime of sensor network. Indeed it should be considered that proposal protocol have close relationship with application. For example the proposal protocol is good for applications such as monitoring temperature changes that have suddenly changes and cannot be used in application that require periodic information reports. Hawaii International Conference on System Sciences (HICSS '00), January [4] A. Manjeshwar and D. P. Agrawal, TEEN: A Protocol for Enhanced Efficiency in Wireless Sensor Networks, in the Proceedings of the 1st International Workshop on Parallel and Distributed Computing Issues in Wireless Networks and Mobile Computing, San Francisco, CA, April [5] Y. Xu, J. Heidemann and D. Estrin, Geography-informed energy conservation for ad hoc routing, in the Proceedings of the 7th Annual ACM/IEEE International Conference on Mobile Computing and Networking (MobiCom 01), Rome, Italy, July [6] R. Kannan, L. Ray, S. S. Iyengar and R. Kalidindi, Max- Min Length-Energy-Constrained Routing in Wireless Sensor Networks, LNCS-Lecture Notes in Computer Science, Springer-Verlag, Vol 2920, (from 1st European Workshop on Wireless Sensor Networks EWSN'2004), January 2004, pp [7] C. Pomalaza-Ráez, "Media Access and Routing Protocols for Power Constrained Ad Hoc Networks", Centre for Wireless Communications University of Oulu and Indiana University - Purdue University, USA, [8] A. Savvides, C. C. Han and M. Srivastava, Dynamic Finegrained localization in Ad-Hoc networks of sensors, Proceedings of the Seventh ACM Annual International Conference on Mobile Computing and Networking (MobiCom), July 2001, pp [9] L. Bajaj, M. takai, R. Ahuja, R. Bagrodia and M. Gerla, "Glomosim: a Scalable Network Simulation Environment", Technical Report , Computer Science Department, UCLA, References [1] W. Heinzelman, J. Kulik and H. Balakrishnan, Adaptive protocols for information dissemination in wireless sensor networks, in the Proceedings of the 5th Annual ACM/IEEE International Conference on Mobile Computing and Networking (MobiCom 99), Seattle, WA, August 1999, pp [2] C. Schurgers and M. B. Srivastava, Energy efficient routing in wireless sensor networks, in the MILCOM Proceedings on Communications for Network-Centric Operations: Creating the Information Force, McLean, VA, [3] W. Heinzelman, A. Chandrakasan and H. Balakrishnan, Energy - Efficient Communication Protocol for Wireless Microsensor Networks, in the Proceedings of the 33rd

126 A Knowledge Management Model to Improve Information Security 105 Yogesh Kumar Mittal 1, Dr Santanu Roy 2 and Dr. Manu Saxena 3 1 Ajay Kumar Garg Engineering College Ghaziabad,Uttar Pradesh, India Research Scholar, Singhania University Jhunjhunu,Rajasthan,India 2 Institute of Management Technology Ghaziabad,Uttar Pradesh, India 3 Human Resource Development Centre Ghaziabad,Uttar Pradesh, India Abstract Information security is an important issue in today s world. Information security management can no more be done by merely a set of hardware and software, rather, it requires a complete endto-end system. Such a system is called Information Security Management System. We have proposed a model to improve Information Security using knowledge management techniques. The model has three modules naming information security knowledge repository module, information security knowledge sharing and dissemination module and information security knowledge Implementation & effectiveness module, first module is to store the information security knowledge in a systematic and easy to use format, second module promote sharing and dissemination of knowledge, third module is responsible to monitor and measure effectiveness of total system respectively. We have allocated knowledge management tools to each module to achieve goals of each module, then we have analyzed relationship between these modules. Keywords: Information Security Management, Information Security, Knowledge Management, Information Security Knowledge Management 1.0 Introduction Managing the growing problem of computer frauds has led researchers and practitioners to emphasize the need to take into account the 'social' aspects of information security(is). In addition, wider organizational issues such as lack of awareness have been associated with computer fraud. Inspecting the domestic and foreign each kind of information security event, it is discovered that behind information security events, those who play the decision role is the human, human's behavior and the information security is closely related, human's unsafe behavior causes accident's primary cause[1]. It requires special focus and participation from all levels of employees with full commitments and responsibilities in establishing such a system and implementing it. In trying to minimize 'opportunities' for computer fraud, managers awareness and knowledge of how an organization information security functions can significantly affect the effectiveness of management of information security. This is because managers can send 'cues' to other employees, which influence how the latter perceive and abide by information security and other policies and procedures in their daily activities [2]. When InfoSec meets Knowledge Management(KM) on process level, security knowledge can be helpful to enhance the effectiveness and maturity of InfoSec[3]. In managerial view, the security problems are invoked because of the short of knowledge to protecting the targets (systems)[4]. In this paper we are proposing an Information Security Knowledge(ISK) Management model. 2.0 Knowledge Management Model to Improve Information Security The proposed model consists of three modules(refer fig. 1). 2.1 ISK Repository Module In this module we will store knowledge related to IS. This will include:

127 106 (a) Standards: Information Security Management System(ISMS) standards like ISO 27001, COBIT and other standards may be stored. (b) Best practices: Industry specific best practices are always evolving with experience. People like to contribute and refer to these practices. (c) Threats and solutions: Information related to current threats and solutions should be stored, which can be easily referred by the users. (b)internal user : Internal users should keep themselves updated with IS related knowledge because they are on the edge of information system. Users may find new problems or ambiguous situation to be discussed with the experts. (c)external users and external experts: Organizations are having extended networks in the form of suppliers and customers, they may also face IS problems. Company s authorized external experts may also take part in sharing the ISK. External experts are having links with the other companies and outer world also, so they can help in sharing. Knowledge regarding coming up IS threats and solutions. 2.3 ISK Implementation & Effectiveness Module In this module we want to ensure that ISK is implemented and the whole exercise of information security knowledge management(iskm) is effective. This module has three components: (a)incentives and recognition to contributors: One of the major problem in application of KM is that experts/knowledgeable persons don t want to share their knowledge, so here we emphasize that incentives recognition and incentives should be given to the contributors of Knowledge. Fig 1 Knowledge Management Model to Improve Information Security 2.2 ISK Sharing and Dissemination Module This module includes sharing ISK between stake holders, disseminate this knowledge to the concerned people and update the ISK repository with the new created knowledge by the users. This module will include: (a)internal experts: Internal information security experts are mainly responsible for any type of security breaches in the organization. They have to address security related issues of the organization and to ensure that the users are understanding and applying this information. (b)incentives and recognition to ISKM team: Overall IS effectiveness should be measured in terms of reduction in IS incidents and better awareness of employees. ISKM team s effort should be recognized and incentives should be given. (c)recognition to users: Users should also be recognized for their active participation, awareness and effectively applying the knowledge available through ISKM activities. Incentives should also be given to the users for their remarkable Knowledge application and contribution for ISKM activities. Above described ISKM activities are to be implemented through the KM tools. Now we will analyse, role of KM tools. 3.0 KM Tools for ISKM Implementation Following KM tools can be used to develop a well structured ISK repository.

128 KM Tools for ISK Repository Module (a)content management system: Content related to standards, best practices, new threats and solutions may be stored in a well structured manner using content management system. (b)taxonomies: Many times, it is difficult to understand the meaning of the technical terms by the common users. Taxonomies is the way of naming technical terms in a natural way or using metaphors and then grouping the information in a convenient way keeping user in view. (c)yellow pages: Yellow pages is like a directory of experts, In which one can search the experts in no. of ways. It may be on the basis of expertise, on the basis of location, on the basis of name of expert etc. (d)e-learning: E-Learning methods like interactive video lectures, quizzes etc can be used by the users sitting anywhere in the world to grasp the basic and advanced knowledge related to IS. This may include knowledge related to standards, best practices, new threats and solutions etc. (e)story telling and narrations: This KM technique can be used to explain concepts like social engineering, virus attacks, disaster and recovery activities etc. (f)enterprise portals: This tool can be used as a gateway for all type of information security knowledge and ISKM activities which can be accessed from any where in the world. 3.2 KM Tools for ISK Sharing and Dissemination Module: Following KM tools can be used for ISK sharing and dissemination. (a)communities of practice(cop): Online CoPs may be formed by stakeholders. Different type of CoPs may be formed on the basis of their expertise, interest and usage. (b)groupware: To disclose knowledge personnel affinity is very important. Group ware develops that affinity using the group dynamics. People become closer using these groups and knowledge sharing become easier. (c) Social network analysis and design: People want to connect socially instead of professionally. Providing opportunities for social networking helps in making social relations. This helps in knowledge sharing. Social network analysis and design will help in designing social networks as per the need and social attributes of the people. (d)seminar/training programs/ workshops: Regular seminars/training programs and workshops can help the users and other stake holders to upgrade their knowledge and skills. (e)mobile devices : Mobile devices can be used to update the security related knowledge immediately and on the move and any where in the world. Like new virus attacks or any disaster condition can be informed immediately. (f)innovation and idea management: Through innovation and idea management system new ideas may be created, nurtured and perfected. The persons contributed that idea may be recognized and rewarded. 3.3 KM Tools for ISK Implementation & Effectiveness Module Implementation of Information Security is the main aim of ISKM. Here we should have following tools. (a)effectiveness metrics: Effectiveness metrics should be engineered. Which may include points for user involvement and implementation, reduction in security incidence, contribution, KM team effectiveness etc. On the basis of this metrics rewards and recognition should be given to the contributors, users and ISKM team. (b)monitoring Mechanism: A monitoring mechanism is required to get the information about contributors, active users & over all information security performance. This can be done by preserving logs of knowledge sharing sessions, security incident logs, feedback from different stake holders. 4.0 Relationship between ISK Repository Module, ISK Sharing & Dissemination and ISK Implementation & Effectiveness Module: All the three modules have to interact with each other and take help from each other to fulfill their objectives. ISK repository will store all types of knowledge available and

129 108 captured in second module. Second module will create new knowledge but may need to access already created knowledge from the first module whenever required. This way second module will enrich the knowledge repository of the first module. Third module is related to implementation and effectiveness, it will refer the captured knowledge for solving day today IS routine work, it will search solution of a particular problems in first module, it may refer second module for newer problems. Both first and second module will receive feed backs from the third module regarding effectiveness of the solutions available in first module and second module. 5.0 Conclusion Proposed model uses KM techniques to effectively educate users regarding information security. This Model describes how different KM tools can be used for different functions of information security knowledge management. All the three modules of this model should be in sync for getting better results from ISKM efforts. This model can be applied for other areas where knowledge is changing with time and updating of stakeholders with that knowledge is important. National/International conferences/journals. His academic and research interest includes IT in Business, Knowledge management, Software Project Management, Enterprise Resource Planning, Software Engineering, Information security and Auditing, Social and Cultural issues. Dr. Santanu Roy is currently serving as a Professor, Operations Management Area, at Institute of Management Technology (IMT), Ghaziabad, India. Dr. Roy had earlier served as a Senior Scientist (Scientist F) in National Institute of Science, Technology and Development Studies (NISTADS), New Delhi. Dr. Santanu Roy has done his Ph.D. in Industrial Engineering and Management from IIT Kharagpur, India and Integrated Master of Science (M.S.) from IIT Delhi. He has more than 26 years of experience in research, consultancy and teaching. Dr. Manu Saxena did B. Sc. in 1977 from, Meerut University, India, M. Sc. in 1979 from University of Roorkee, Roorkee, India Ph. D. from University of Roorkee, Roorkee, India in Operational Research in He published 19 papers in national/international conferences and journals. He supervised 13 dissertations of post graduation level. References [1]Behavioral science-based information security research, Yang yue jiang Yu yong xia, 2009, First International Workshop on Education Technology and Computer Science IEEE [2]Knowledge management within information security: the case of Barings Bank, Shalini Kesar, International Journal of Business Information Systems 2008 Vol. 3, No.6 pp [3]Knowledge-Centric Information Security, Walter S. L. Fung, Richard Y. K. Fung, 2008 International Conference on Security Technology, IEEE [4]Knowledge Management Tools & Techniques, Practitioners & Experts Evaluate KM Solutions, Madan Mohan Rao, 2005 Elsevier [5]Knowledge Management in Theory and Practice, Kimiz Dalkir,,2005, Elsevier Inc. Yogesh Kumar Mittal did B.Tech. from Maulana Ajad College of Technology, Bhopal, India (Now MANIT, Bhopal) in 1987 than M.Tech. in Computer Science and Technology from University of Roorkee, Roorkee, India (Now IIT Roorkee) in He also did PGDBM from IMT, Ghaziabad, India in He qualified prestigious CISA (Certified Information System Auditor) exam in He has around 21 years of experience in industry and academia. He has worked as Consultant, Information System Auditor, General Manager and Chief Executive Officer before joining the teaching profession. He published 10 papers in

130 109 TH*: Scalable Distributed Trie Hashing ARIDJ MOHAMED 1,2,ZEGOUR DJAMEL EDINNE 1 1 Institut National d'informatique Alger, Alegria 2 University Hassibe benbouli Department, University Name, Company Chlef, Algeria Abstract In today s world of computers, dealing with huge amounts of data is not unusual. The need to distribute this data in order to increase its availability and increase the performance of accessing it is more urgent than ever. For these reasons it is necessary to develop scalable distributed data structures. In this paper we propose a TH* distributed variant of the Trie Hashing data structure. First we propose Thsw new version of TH without node Nil in digital tree (trie), then this version will be adapted to multicomputer environment. The simulation results reveal that TH* is scalable in the sense that it grows gracefully, one bucket at a time, to a large number of servers, also TH* offers a good storage space utilization and high query efficiency special for ordering operations. Keywords: Trie hashing, Distributed hashing, SDDS, multicomputer, distributed systems. 1. Introduction A multi-computer consists of set of workstations and PCs interconnected by a high speed network such as the Ethernet and TMT. It is well known that multi-computers offer best price-performance ratio; thus offering some new perspectives to high performances applications [11, 13]. In order to achieve these performances, a new class of data structures has been proposed. It is called Scalable Distributed Data Structures (SDDS) [3,4,5,9,8,10,12, 14,16] and is based on the client/server architecture. This new structure supports the parallel processing that does not require the central processing of the addresses. Data is typically stored in the distributed main memory (DRAM). An SDDS may easily handle very large files and their access is achieved in a fraction of the disk access time. An SDDS propagates to new sites through splitting when sites are saturated. The splitting process is transparent to the applications. All SDDSs support key searches where some supports range and/or multi key searches. However, every client has its own file picture. The updating of the file structure are not sent to the clients in a synchronous manner. A client can make addressing error when using its own file picture. Each server verifies the address of the received request and is routed to another server if an address error is detected. The server that processes the requests then sends an adjusted message to the client that made the addressing error. This message is called Image Adjustment Message (IAM).The IAM allows the client to adjust its image to avoid making the same error again. However, its image is not necessarily globally exact. Related work LH*[8] is the first proposed SDDS. It is based on the linear hashing technique [17]. It achieves good performance for single-key operations, but range searches are not performed efficiently. To rectify this, order preserving structures was proposed. Among order preserving SDDSs, we recall RP* family [18],BDST[19] and ADST[20]. Rp*, based on the B+-tree technique and BDST, based on balanced binary search tree achieve good performances for range searches and a reasonably low,but not constant worst-case cost for single key operations. ADST (Aggregation in Distributed Search Tree) obtaining a constant single-key query cost, like LH*, and an optimal cost for range queries, like RP* and DRT* but a logarithmic cost for insert queries producing a split. All these techniques (except RP*N) use an index witch is stored at the servers and/or clients RAMs. This index may take up large space in the RAM. In this paper, we propose a new SDDS, called TH*, that adapts the TH to distributed environments. The remainder of the paper is organized as follows. In section 2 we review Trie Hashing [6] and in Section 3 a

131 110 new TH version which eliminates nil nodes is discussed. The principles and algorithms of the new SDDS TH* are introduced in Section 4 and section 5 discusses the results of an evaluation of TH* using simulation. We conclude and summarize the paper in Section Trie hashing Trie hashing is one of the most powerful access methods used for dynamic and ordered files. The TH file is a set of records identified by primary keys. Keys consist of a finite ordered sequence of characters. Records are stored in buckets and each bucket may contain a fixed or a variable number of records (Fig.1.b). The file is addressed thought the consisting of the nodes where each node is either a leaf or an internal.(fig.1.a). An internal node contains a pair of values (d,i) where d is a character and i is a number representing the position of the character d in the key. However the leaf node contains the bucket address or the value Nil indicating that no buckets are associated with the leaf. 3. THwn: trie hashing without Nil nodes Given that in the trie the Nil nodes appear when more than one digit is necessary for the split string, Litwin s method solves the problem as follows: Let C= d 0 d 1 d 2... d n be the split string necessary to split a bucket m into the bucket m the resulting trie will have the form described in Fig Proposition: To eliminate the nil nodes we propose that they will be replaced by the address of the new allocated bucket (m ).In this case the trie will have the following form given in Fig 3. Fig.1 : TH Principle The insertion operation may cause the expansion of the file and the trie; however the deletion operation may cause a contraction. All the algorithms related to TH are taken from [6]. The load factor of the file is about 70% for random insertions and of 60% to 70% for ascending insertions. The key search takes at most one disk access and insertion takes 2 or 3 disk access. The behaviour and performance analysis of TH can be found in[1,2,7,15]. This proposition gives a new trie hashing scheme we named THwn and which stands for Trie Hashing Without Nil nodes. 3.2 Example : To illustrate the principle of THwn we insert the following keys: abmf, abnm, acnm, aczm, aczh in the file and we assume that the capacity of a bucket is equal to 4. The insertion of the abmf, abnm, acnm, aczm, is done in bucket 0, the insertion of the aczh causes a collision; in this case the splitting string is acn. The trie after splitting is show in Fig.4.

132 111 Our proposal is justified by the behavioural analysis of the bucket range. By observing the example given in Fig4,we note that initially all the keys will be inserted in the bucket because the range of this bucket is [ _, ] (where _ represents the smallest digit and the greatest digit). After the proposed splitting we obtain the configuration given in Fig 5.b. However the application of Litwin s algorithm gives the configuration shown Fig 5.a. 3.3 The THwn Algorithms The algorithm for searching and inserting remain similar to those proposed in [6]. However for bucket splitting we propose the following algorithm given in Fig6 and its illustration using an example is given in Fig Concepts As previously mentioned, TH* is based on the client/server architecture. Each client contains a partial trie which represents the client image of the distributed file and is being updated gradually until we obtain the real trie. Any client can enter the system with an empty trie and each server has a Bucket containing the records of the file, the trie and an interval [Min, Max ] where Min is the smallest value the server can contain and Max the largest value. Initially the system contains only the server 0 with an empty bucket. Its interval is ] _...,.. ] as previously introduced and with an empty trie (leaf with value 0). The file expands by splitting due to collisions. At each collision there is a distribution of the keys (splitting of overflowed server) to another server. The number of servers is conceptually infinite and each server can be determined in a static or dynamic way. When an addressing error occurs, part of the server trie is transferred to the client for updating its trie 4.2 Example of TH* File To illustrate our method we consider a system composed of 4 clients and several servers. We suppose that the server s capacity is 4. The following list of insertions is represented by a set of pairs (Client_Number, keys). (1 js), (1 hw), (3, c), (2, gwmr), (3, g), (2, km), (4, zur), (1,ewg), (3, lewhv), (2, nrq), (3, mf), (4, pem), (4, rl), (2, bqyg), (3, v), (1, j), (2, qcm), (4, czxav), (2, lhgd), (3, z), (1, lrz), (3, kiyfg), (4, pbtpr), (3, hpqtp), (4, h). Fig 8 shows the final state of the system after 25 insertions 4. Distributed trie hashing (TH *) In this section we propose the extension of THwn to a distributed environment, the obtained schema is named TH* Fig.8 : TH* file example.

133 4.3 TH* algorithms addressing : Transformation key --> server address To realise an operation on a TH* file, the client calculates the address of a suitable server by traversing its trie which represents its image on the TH * file. When the server receives a client request, it checks if the key is in its interval. Two cases can occur: Case 1 The key is in the server interval: the server treats the request and sends back an Image Adjustment Message (IAM) to the client if an addressing error had being detected. The client uses the IAM to update its image. Case2 There is an addressing error: the server recalculates the address using its cached file image (trie) and forwards the message to another server as illustrated in Fig Step 1 : constructing the IAM at server using the following algorithm: in this step the server determines the sub trie to send to the client in order to correct its false image (trie). This algorithm can be as follows: Let Cm client : be the maximum key attributed to the concerning server by the client trie.(sent in the client query) 1. [determination of server maximum key] Determine the Cm server the maximum server key by traversing the local trie. 2. [determination of common prefix] let prefix=cm server - Cm client 3. [construction of IAM] if prefix=cm server then IAM= server trie Else IAM= the sub trie of server giving the prefix string 4. [send IAM (sub trie) to the client] Step 2: updating client trie: 1. [receive the IAM from server] Let S the sub trie contained in the IAM 2. [trie updating] Let P be the pointer of the leaf contains m ( resulting from search algorithm) Let m be the server address in the latest leaf of S Replace the leaf m by the S (attach S to the trie) While forward(p) = m do Replace m by m P forward(p). Fig 10 : Adjusting client image algorithm Adjusting client image In the case of an addressing error by the client, the last server participating in the forwarding process sends back the adjustment image message to the client. The client then updates its trie using the algorithm given in Fig 10 which is illustrated by the examples given in Fig 11 and Fig 12. Examples:

134 113 let B be the ordered sequence of (b+1) keys to split, including the new key and Seq the spliting string 1. [trie updating] use the algorithm A1 of section 3.4 to update server m trie. 2.[interval updating] let ]Am,Bm] be the initial interval of server. the new interval of m is: ]Am,Seq+ ] 3.[allocation of new server] allocate new server m and initialise it as flows: trie is empty (one leaf with the m address) interval = ]Seq+,Bm] 4. [keys redistributing] - m bucket contains the keys Seq - m buket contains the keys> seq Insertion The insertion requires in a first instance the check if the key exists using the search algorithm. If the key does not exist in the file, it will be inserted in the appropriate server. However, if this server is full the splitting algorithm is then used. Fig. 13 summarizes the principle of an insertion operation in a SDDS TH* Fig: 14 Sever splitting algorithm range query The range query operation consists on listing all the keys in a given interval. Since the TH preserves the keys order (the trie keeps the order of the keys), this operation is be easy to perform. The general algorithm of this operation is realised in two steps. First the client determines all the servers that could contain the keys, and then the servers process the request. The general algorithm of this operation is as follows :( Fig 15.a and Fig 15.b) At level of client If (kmin is in server ) then If (kmax Cm) then Return all the keys which are in [Kmin, Kmax] Stop<--true ; Else Return all the keys which are in [Kmin, Cm]. Stop<--false ; Else Addressing error: send the IAM to the client Server splitting Fig 15.a : range query algorithm at level of client The splitting algorithm is called by the overflowing server. It allocates a new server and redistributes the keys. The algorithm is given in Fig14

135 114 At level of server let [kmin,kmax] be the range query interval 1. Search kmin in the local trie : let m be the result, and Cm its maximum keys of m. 2. Send request to the server m 3. at the reception of servers reply one of they tree case is executed : Case1 : error address : use the adjustment algorithm to updating the client trie Case 2 : Stop=true : server continuing Kmax is attended processes is stooped Case 3 : Stop=false : use the server trie to find the next server. Comment: Observation of the figure (Fig 16) shows that: - TH* offers good storage space utilization, witch it is about (65 % to 93%). - The server capacity does not influence the behavior of load factor. b) Variation of number of servers: Under the same hypotheses, we have noticed variations of the servers splitting (Fig 17). Fig 15.b : range query algorithm at level of server 5. Simulation and test results: 5.1 Development environment We have implemented the SDDSs TH* on a multicomputer composed of 4PCs executing the LINUX system (Mandrake 8) and connected by a high seeped network. The PCs are Pentium V 900 MHz with 128 Mo of RAM. Every machine can be client and/or server 5.2 Some results a) Load factor We have inserted 1,000,000 keys in a TH* files with different server capacity (b=50,b=100,b=500,b=1000). The load factor variations can be summarized in the following graph (Fig 16) This experiment showed that the server capacity does not affect the behavior of load factor which is giving by figure (Fig.16).. Comment: The figure (Fig 17) shows that the number of the servers splitting increases linearly with the number of the inserted keys c) variation of search and insertion average time The study of the two parameters is very important because it makes the method either scalable or not. The curve of figure (fig.18) represents a research and insertion average time. Comment: The two curves are practically linear. This implies that the research and insertion average time does not increase with the number of keys, which makes the method very scalable. 1. Conclusion This paper presented an adaptation of Trie hashing to distributed environments. The method is based on the client/server architecture and respects the SDDS properties. The obtained schema is called TH* and its behavioural analysis shows that it is efficient and scalable during insertions and retrievals, with

136 115 performance close to optimal. It also offers good storage space utilization. Since TH* is an efficient, scalable, distributed data structure, it provides a new method to be used in applications such as next generation databases, distributed data mining, distributed data warehouse, distributed dictionaries, bulletin boards, etc Acknowledgments The authors would like to thank Litwin Witold and the members of CERIA research group for taking the time discuss the ideas presented here. References [1] M.aridj, hachage digitale compact multidimesionnel avec expansion partielle, these de magister -INI [2] M.aridj,D.E Zegour, A new multi-attributes access method for voluminous files, SPECTS 2005 Summer Simulation Multiconference July 24 28, 2005 Philadelphia,Pennsylvania, USA. [3] M.Aridj,LH*TH: New fast Scalable Distributed Data Structures SDDS NET Technologies 2005 International Conference May 29 June 1, 2006 University of West Bohemia, Plzen, Czech Republic. [4] Devine, R. Design and Implementation of DDH: Distributed Dynamic Hashing. Intl. Conf. On Foundations of Data Organizations, FODO-93. Lecture Notes in Comp. Sc., Springer-Verlag (publ.), Oct [5] Karlsson, J. Litwin, W., Risch, T. LH*lh: A Scalable High Performance Data Structure for Switched Multicomputers. Intl. Conf. on Extending Database Technology, EDBT-96, vignon, March [6] Litwin,W : Trie Haching. Proc.ACM. SIGMOD 81,pp [7] W.Litwin «Trie hashing : Further Properties and Performances» Int.Conf. on Foundation of Data Organization. Kyoto, May 1985, Plenum press [8] W.litwin Neimat, M-A., Schneider, D, LH* : Linear Hashing for Distributed Files. ACM-SIGMOD Intl. Conf. On Management of Data, [9] W.litwin Neimat, M-A., Schneider, D, RP* A Family of Order-Preserving Scalable Distributed Data Structures, Proceedings of the 20th VLDB Conference Santiago, Chile, 1994 [10] Litwin, W., Neimat, A.M, High Availability LH* Schemes with Mirroring, Intl. Conf on Cooperating systems, Brussels, IEEE Press 1996 [11] Tanenbaum, A., S, Distributed Operating Systems. Prentice Hall, 1995, 601. [12] Tung, S, Zha, H, Kefe, T, Concurrent Scalable Distributed Data Structures, Proceedings of the ISCA International Conference on Parallel and Distributed Computing Systems, pp , Dijon, France, September,1996. Edited by K. Yetongnon and S. Harini [13] Ullman, J. New Frontiers in Database System Research. Future Tendencies in Computer Science,Control, and Applied Mathematics. Lecture Notes in Computer Science 653, Springer-Verlag, A. Bensoussan, J. P. Verjus, ed [14] Vingralek, R., Breitbart, Y., Weikum, G, Distributed File Organization with SalableCost/Performance. ACM- SIGMOD Int. Conf. On Management of Data, [15] D.E.Zegour,W,Litwin,G.Levy,Multilevel trie hashing int. conf on VLDB venise Italy [16] D.E.Zegour, Scalable distributed compact trie hashing, Elsevier information and software technology 46 P November [17] Litwin, W. Linear Hashing : a new tool for file and tables addressing. Reprinted from VLDB-80 in reading in database 2-nd ed. Morgan Kaufmann Publishers, Inc.,1994. Stonebraker, M.(Ed.).1999 [18] Litwin W, Neimat M-A, Schneider D : RP* : A Family of Order-Preserving Scalable Distributed Data Structures. Proceedings of the 20th VLDB Conference Santiago, Chile, [19] Pasquale A.Di, Nardelli E: Fully Dynamic Balanced and Distributed SearchTrees with Logarithmic Costs, Workshop on Distributed Data and Structures (WDAS 99), Princeton, NJ, May [20] Pasquale A.Di, Nardelli E: ADST: An Order Preserving Scalable Distributed Data structure with Constant Access Costs L. Pacholski and P. Ruˇziˇcka (Eds.): SOFSEM 2001, LNCS 2234, pp , [21] Pasquale A.Di, Nardelli E: Scalable Distributed Data Structures: a Survey. In 3rd International Workshop on Distributed Data and Structures (WDAS'00), pages , L'Aquila, Italy, June Aridj Mohamed is assistant professor at university Hassiba Benboali chlef Algeria since His area of interest include Software Engineering, distributed System, Databases, access methods and hashing Zegour djamel edinne is professor at institute national d informatique Algeria. His area of interest includes Software Engineering, System Analysis and Design, Databases, distributed systems, access methods and Object Oriented Technologies.

137 116 Attacks in WEB Based Embedded Applications Yaashuwanth.C, Research scholar Dept. of Electrical and Electronics Engineering, Anna University Chennai, Chennai , Tamilnadu, India. Dr. R. Ramesh, Assistant Professor, Dept. of Electrical and Electronics Engineering, Anna University Chennai, Chennai , Tamilnadu, India. Abstract This paper deals with the issues related to embedded applications when they are implemented in internet. There are various attacks in embedded systems when implemented in the internet. These attacks have a negligible effect in the operating system which results in the decrease in the system performance. But in embedded system case, it has life and death consequence attached to it. Many of these embedded systems work in hazardous environment where a system failure results to catastrophic effects. Here a study of various attacks are discussed and a new architecture has been proposed to secure Web based embedded systems from the attacks. Keywords: Embedded systems, Web, Internet, attacks, secured layer 1. Introduction There are a number of security algorithms which prevent attacks when Web services are implemented. In general purpose operating systems, there are hardly a handful of algorithm which prevents Web attacks in Embedded systems. This paper gives a general study of various attacks that happens in Web based embedded systems. Dong Haung [1] proposed a new ontology for representing security constraints as policy and a semantic policy framework for the management of the policies. The growth of internet has accompanied the growth of e services which resulted in increasing attacks on them by malicious individuals. The authors [4] highlighted the need of security. The conceptions about security of Web services and Degree of Safety for Web Services (WS-Dos), and the duration of Web service execution time, are introduced in the paper. In addition, a securing logical hierarchical structure for Web Service application based on an extended Web Services security architecture model with five elemental objects, such as resources, services, roles, protocols and methods object is analyzed Moreover, an integrated security solution has been developed and the results of application showed that the solution builds a confidence and authentication security environment for all roles in the process of dynamic B2B trade [8]. The authors [5] sketched the MAWeS architecture, illustrating how to use it to optimize the performance of a typical compound Web Services application while at the same time guaranteeing that a set of security requirements, expressed by a security policy, are met.chen el al [6] aimed at clarifying security concern by conducting a quantitative performance evaluation of WSS overhead. Based on the evaluation, an extension of the existing Web services performance model is made by taking the extra WSS that overheads into account. The extended performance model is validated on different environments with different messages sizes and WSS security policies. Micheal lesk et al [7] explained Web Service security in a federated environment that describes how the Web services are implemented and are secured in a Web environment Kevin [2] described the objective of improving Internet messaging by redesigning it as a family of Web services, an approach that is call WS . This paper illustrated architecture and describe some applications. Since increased flexibility often mitigates against security and performance, here the steps for proving security properties and measuring the performance of the system with its security operations are focussed. The authors proposed [3] an agent based policy aware framework for Web Services security. In this framework, a policy language called ReiT which is a declarative language based on the rules and ontology is introduced. The non-structural knowledge is represented by rules and the structural temporal knowledge is represented by ontology Moreover, the authors propose a mixed reasoning mechanism to evaluate the ReiT policy. The access control policy including the context of the user and Web Services is evaluated by the reason or in addition, policy aware BOID agent to authorize the access control of the Web Service is presented. And the authors implemented the policy aware of BOID agent by extending the JADE platform. 2. Study of various attacks in Real Time Embedded Systems 2.1 Denial of service and distributed denial of service attacks. A denial-of-service attack (DoS attack) or distributed denial-of-service attack (DDoS attack) is an attempt to make a computer resource unavailable to its intended users. Although the means to carry out, motives for, and targets of a DoS attack may vary, it generally consists of the concerted efforts of a person or people to prevent an Internet site or service from functioning efficiently or at all, temporarily or indefinitely. Propagators of DoS attacks typically target sites or services hosted on high-profile

138 117 Web servers such as banks, credit card payment gateways, and even root nameservers. The term is generally used with regards to computer networks, but is not limited to this field, For example, it is also used in reference to CPU resource management. One common method of attack involves saturating the target machine with external communications requests, such that it cannot respond to legitimate traffic, or responds so slowly as to be rendered effectively unavailable. In general terms, DoS attacks are implemented by either forcing the targeted computer(s) to reset, or consuming its resources so that it can no longer provide its intended service or obstructing the communication media between the intended users and the victim so that they can no longer communicate adequately.one step ahead, DDos does is capable of doing more harm. With this, attacker can use the victims system to infect other connected systems or send a spam. Attacker can find a weakness in the system and can inject a malware or software which can be remotely used by using this, now attacker can make the server a slave and send spams or get access to files using its permission. Thousands of system can be targeted from a single point. DOS and DDOS can also happen in embedded systems since the malicious hacker can gain access to the embedded Web server and use all the server resources such as limited bandwidth which in turn leads to denial of service for legitimate embedded client from accessing the service 2.2 Threat from Key Logging. Keystroke logging (often called keylogging) is an action of tracking (or logging) the keys struck on a keyboard, typically in a covert manner so that the person using the keyboard is unaware that their actions are being monitored. There are numerous keylogging methods, ranging from hardware and software-based approaches to electromagnetic and acoustic analysis. The users in the embedded Web client must ensure that their keystroke logging must not be tracked by an imposter. The snooper can gain entry if he is able to track the key logging of the end user. 2.3 IP Spoofing. IP address spoofing or IP spoofing refers to the creation of Internet Protocol (IP) packets with a forged source IP address, called spoofing, with the purpose of concealing the identity of the sender or impersonating another computing system or an embedded device. 2.4 Buffer Overflow A buffer overflow, or buffer overrun, is an anomaly where a program, while writing data to a buffer, overruns the buffer's boundary and overwrites adjacent memory. This may result in erratic program behavior, including memory access errors, incorrect results, program termination (a crash), or a breach of system security. Buffer overflows can be triggered by inputs that are designed to execute code, or alter the way the program operates. They are thus the basis of many software vulnerabilities and can be maliciously exploited. Bounds checking can prevent buffer overflows. Programming languages commonly associated with buffer overflows include C and C++, which provide no built-in protection against accessing or overwriting data in any part of memory and do not automatically check that data written to an array which (the built-in buffer type) is within the boundaries of that array In embedded system this attack poses a greater threat since embedded systems will only allocate a small amount of memory. So any small type of this attack may lead to overflow of buffer at end user side and hence the system will crash 2.5 Format String Attack. Format string attacks are a class of software vulnerability. Format string attacks can be used to crash a program or to execute harmful code. The problem stems from the use of unfiltered user input as the format string parameter in certain C functions that perform formatting, such as printf(). A malicious user may use the %s and %x format tokens, among others, to print data from the stack or possibly other locations in memory. One may also write arbitrary data to arbitrary locations using the %n format token, which commands printf() and similar functions to write the number of bytes formatted to an address stored on the stack This type of attack is also common in embedded systems which causes the embedded systems to crash 2.6 SQL Injection Attack. SQL injection is a code injection technique that exploits a security vulnerability occurring in the database layer of an application. The vulnerability is present when user input is either incorrectly filtered for string literal escape characters embedded in SQL statements or user input is not strongly typed and thereby unexpectedly executed. It is an instance of a more general class of vulnerabilities that can occur whenever one programs or scripts the language that is embedded inside another. SQL injection attacks are also known as SQL insertion attacks. 2.7 Cross site scripting. Cross-site scripting (XSS) is a type of computer security vulnerability typically found in Web applications that enables malicious attackers to inject client-side script into Web pages viewed by other users. An exploited cross-site scripting vulnerability can be used by attackers to bypass access controls such as the same origin policy. Cross-site scripting carried out on Websites were roughly 80% of all security vulnerabilities documented by Symantec as in

139 Their impact may range from a petty nuisance to a significant security risk, depending on the sensitivity of the data handled by the vulnerable site, and the nature of any security mitigations implemented by the site's owner. 2.8 Virus and worms. A virus is a program that can copy itself and infect a computer or any embedded device. The term "virus" is also commonly but erroneously used to refer to other types of malware, including but not limited to adware and spyware programs that do not have the reproductive ability. A true virus can spread from one computer to another (in some form of executable code) when its host is taken to the target computer; for instance a user can sent it over a network or the Internet, or carried it on a removable medium such as a floppy disk, CD, DVD, or USB drive. Viruses can increase their chances of spreading to other computers by infecting files on a network file system or a file system that is accessed by another computer. As stated above, the term "virus" is sometimes used as a catch-all phrase to include all types of malware, even those that do not have the reproductive ability. Malware includes computer viruses, computer worms, Trojan horses, most rootkits, spyware, dishonest adware and other malicious and unwanted software, including true viruses. Viruses are sometimes confused with worms and Trojan horses, which are technically different. A worm can exploit security vulnerabilities to spread itself automatically to other computers through networks, while a Trojan horse is a program that appears harmless but hides malicious functions. Worms and Trojan horses, like viruses, may harm a computer system data or performance. Some viruses and other malware have symptoms noticeable to the computer user, but many are surreptitious or simply do nothing to call attention to themselves. Some viruses do nothing beyond reproducing themselves. 2.9 Password cracking Password cracking is the process of recovering passwords from data that has been stored in or transmitted by a computer system and embedded client systems. A common approach is to repeatedly try guesses for the password. The purpose of password cracking might be to help a user recover a forgotten password (though installing an entirely new password is less of a security risk, but involves system administration privileges), to gain unauthorized access to a system, or as a preventive measure by system administrators to check for easily crackable passwords. 3. Proposed Architecture When real time applications are implemented with Web services they are subjected to various attacks which has been discussed. So there is a need to develop a proposed architecture to counter these attacks. The proposed architecture solves the drawbacks of all the issues related to Web based embedded systems. The proposed architecture identifies the source of attacks. This is accomplished by deploying our defense systems in our distributed routers in order to examine incoming messages and place the headers. By this way, we can examine the message header and traceback the originator and prevent further attacks. Embedded Web Client Embedded Web server Embedded Web client Request Request Response Response Request Intranet or Internet Request Secured layer Embedded Web client Response Response Fig. 3.1 Proposed architecture

140 Conclusion This paper finally tune all the attacks related to embedded systems these attacks are related to denial of service to the Web client. Thus also a proposed architecture which is developed and implemented to prevent these type of attacks. References [1] Dong Huang Scematic Description of Web Service Security Constraints Proceedings of the Second IEEE International Symposium on Service-Oriented System Engineering IEEE number= [2] Kevin D. Lux, Michael J. May, Nayan L. Bhattad, and Carl A. Gunter WS Secure Internet Messaging Based on Web Services Proceedings of the IEEE International Conference on Web Services IEEE [3] Jlan-xin li,bin li,liang li and Tong sheng che An Agentbased Policy Aware Framework for Web Services Security IFIP International Conference on Network and Parallel Computing Workshops IEEE [3] George Yee and Larry Korba Negotiated Security Policies for E-Services and Web Services Proceedings of the IEEE International Conference on Web Services (I IEEE [4] Massimiliano Rak, Valentina Casola, Nicola Mazzoccca Emilio Pasquale Mancini and Umberto Villano Optimizing Secure Web Services with MAWeS: a Case Study pdf [5] Shiping Chen1, John Zic, Kezhe Tang, and David Lev Performance Evaluation and Modeling of Web Services Security International Conference on Web Services IEEE [6] Michael Lesk, Martin R. Stytz and Roland L. Trope providing Web service security in a federated environment International conference on Privacy and Security IEEE pdf [7] Yuan wo,, Bo-qin feng', Jin-Cang and Zun-Chao li SX- RSRPM: A Security Integrated Model for Web Services Proceedings of the Third International Conference on Machine Learning and Cybernetics IEEE About the authors Mr. C.Yaashuwanth completed his B.Tech. degree in Information technology at BSA CRESCENT Engineering College Chennai, He completed M.E. in Embedded System Technologies at VEL TECH Engineering college Chennai. He is currently Pursuing his Ph.D program under Dr. R.Ramesh Dr. R. Ramesh pursued his B.E. degree in Electrical and Electronics Engineering at University of Madras, Chennai, and completed his M.E. degree in Power Systems Engineering at Annamalai University, Chidambaram. He received his Ph.D. degree from Anna University, Chennai and has been a faculty of Electrical and Electronics Engineering Department of College of Engineering, Guindy, Anna University, Chennai since His area of interest are Real-time Distributed Embedded Control On-line power system analysis and Web services.

141 A Tool for Qualitative Causal Reasoning On Complex Systems 120 Tahar Guerram 1, Ramdane Maamri 2 and Zaidi Sahnoun 2 1 Departement of Computer Science University of Oum El Bouaghi, ALGERIA 2 LIRE Laboratory, Departement of computer Science University Mentouri of Constantine, ALGERIA Abstract A cognitive map, also called a mental map, is a representation and reasoning model on causal knowledge. It is a directed, labeled and cyclic graph whose nodes represent causes or effects and whose arcs represent causal relations between these nodes such as increases, decreases, supports, and disadvantages. A cognitive map represents beliefs (knowledge) which we lay out about a given domain of discourse and is useful as a means of decision making support. There are several types of cognitive maps but the most used are fuzzy cognitive maps. This last treat the cases of existence and no nexistence of relations between nodes but does not deal with the case when these relations are indeterminate. Neutrosophic cognitive maps proposed by F. Smarandache [1] make it possible to take into account these indetermination and thus constitute an extension of fuzzy cognitive maps. This article tries to propose a modeling and reasoning tool for complex systems based on neutrosophic cognitive maps. In order to be able to evaluate our work, we applied our tool to a medical case which is the viral infection biological process. Keywords: qualitative reasoning, fuzzy cognitive maps, neutrosophic cognitive maps, causal reasoning. Complex systems 1. Introduction Causal reasoning is a cognitive activity that a human being practices in its everyday life to try to provide explanations to physical, social, economic and ecologic phenomena or facts. It is a qualitative and a common sense reasoning approach. In 1976, Robert Axelrod [2], researcher in political sciences, introduced the concept of a cognitive map as being a directed, labeled and a cyclic graph, whose basic relations set is {+, -, 0}, respectively representing the relations of causality or influence "increases", "decreases" and "no effect". Axelrod s cognitive maps were used to represent, predict and make decisions relating to the socio political and socio economic fields, which are considered as complex systems. In 1986, Kosko [3] proposed an extension of Axelrod s cognitive maps by introducing fuzzy logic and named its maps by fuzzy cognitive maps where the causal relations are graduated in the interval [- 1 1]. The fuzzy cognitive maps were mainly used to model complex supervision systems [4]. In 2003, F. Smarandache [1] proposed the neutrosophic cognitive maps making it possible to mitigate the limitation of fuzzy cognitive maps which cannot represent the indeterminate relations between variables. The capability of neutrosophic cognitive maps to represent indetermination makes it possible to apprehend the complexity of the systems and thus to elucidate and predict their behaviors in the absence of complete information. Our paper will be structured in the following way. Section 2 presents related work in the field of modeling biological complex systems using cognitive maps. The section 3 presents neutrosophic cognitive maps. Section 4 will be reserved to describe the domain of discourse in fact the viral infection then to build the corresponding neutrosophic cognitive map by the recourse to a certain number of virologists. Reasoning on this cognitive map making it possible to draw the typical behaviors of the viral infection process. Section 5 is devoted to describe implementation aspects. Our objective is to automate the construction and the reasoning on the neutrosophic cognitive maps independently of the field of expertise. In order to be able to evaluate our work and to discuss the results, section 6 shows us how our developed tool is applied to the study of the viral infection process. Finally section 7, will enable us to conclude our work and to give some perspectives. 2. Related work Fuzzy cognitive maps [3] were used in the medical domain [7][8][9] as a tool of causal reasoning, procedure often complex because multi sources data is often vague, conflicting, missing and no easy to interpret. There is a few works done on cognitive maps to model and reason about biological processes and we have only found two significant papers written by Hailin and Chenxi [10] and by Papageoriou et al. [11].

142 121 The first paper [10] introduced the qualitative reasoning technology to model and to analyze the virus penetration in the animal cell. The authors distinguish three classes of factors (See table 1) which influence the viral behavior which are: 1) Factors relating to the virus, including some situations about the virus such as its activity, its size and its structure. 2) Factors relating to the environment including the temperature, the ph and the presence of positive ions. 3) Factors relating to the cellule including permissiveness of the cellular membrane and its enzyme countenance. Each factor is affected by a degree of influence as shown in the Table 1. According to the authors of this paper, the process of entry of the virus to the cell passes by three stages: 1) Attachment of the virus to the cellular membrane, 2) Penetration of the virus inside the cell, 3) Release of the genetic inheritance of the virus inside the cell. Based on this information, a qualitative model, based on the method of T.L. Saaty proposed in 1970, was built and tested but remains an approximate model prone to improvements being the subject of future work [10]. In the second work suggested by Papageogiou and his/her colleagues [11], authors proposed a model based on the fuzzy cognitive maps to diagnose the severity of the pulmonary infection being able to have four fuzzy values: weak, moderated, average and severe. For that they connect factors having milked at the symptoms, the physical examinations, the biomedical and biochemical analyses and other factors relating to resistance to antibiotics. In total, thirty four (34) concepts are identified and connected thanks to the assistance of the experts invited to define the degree of influence between these concepts. (See Figure 1). The model builds was tested on real medical data and the simulation results showed that the reliability of the decision support system used, is dependent on the availability of sufficient information [11]. Figure 1. Fuzzy cognitive map of the viral infection according to [10] Factors Influence degree The individual factors 1 Virus activity 9 2 The structure of 8 virus attachment protein The environment factors 3 temperature 5 4 PH value 5 5 Positive ion 5 The Cell factors 6 Receptor site 7 7 Enzyme 5 8 The structure of 6 cellular skeleton Our work consists of modeling and simulating the viral infection process by using not only the direct factors between the virus and the cell but also other factors having milked with the environment and/or the lived socio economic situation of the patient. Moreover, the model used, which is the neutrosophic cognitive map [1], is an extension of the fuzzy cognitive map model allowing to take into account of the indeterminism, significant characteristic of biological processes and of complex systems in general. Table 1: factors influencing the viral Infection according to [10 ]. 3. Neutrosophic cognitive maps, representation of causal knowledge and reasoning In this section, we will show how to represent knowledge related to a field of expertise by a Neutrosophic Cognitive Map then to explain the manner of reasoning on this latter in order to draw typical behaviors from the dynamic system which it represents. A Neutrosophic Cognitive Map (N.C.M) is an extension of a Fuzzy Cognitive Map (F.C.M). The latter is a cyclic

143 122 graph directed whose nodes represent the variables of the field of speech and whose arcs define causal relations between each pair of variables. These causal relations take their values in the interval [ ]. N.C.M take their values of causalities in the same preceding interval but can also be indeterminate. Cognitive maps whose values of causalities belong to the set {-1,0,1,?} is known as Simple Neutrosophic Cognitive Maps [1]. A state vector associated with a cognitive map is given by the vector: S = (s 1,s 2,s 3,s N, ), where the state of node i could be represented by values 1, 0 and? whose significances are respectively, active, inactive or indeterminate. The reasoning carried out on a cognitive map makes it possible to draw all the typical behaviors from this map, i.e. the typical behaviors of the dynamic system that this cognitive map represents. Each behavior of the dynamic system is obtained by the stimulation of the adjacency matrix of the cognitive map by a suitable state vector via an iterative multiplication until we fall on a state vector which is already obtained in the preceding iterations. Before applying the K ieme iteration, a thresholding function is applied to the state vector of the (K-1) eme iteration defined as follows [1]:?, 0 0? 1, 0 0 0, 0 A typical behavior represents an equilibrium state of the cognitive map and it is not other than a fixed point or a limit cycle. In order to illustrate this, we give the following example. c8 C1 C2 C3 Example [1]: The simple neutrosophic cognitive map of Figure 2.1 represents causal knowledge relating to the phenomenon of s hacking by students. The arcs in dotted lines represent indeterminate causal relations. Concepts of this cognitive map are: C1 (Curiosity); C2 (Professional rivalry); C3 (Jealousy/ enmity); C4 (Sexual satisfaction) ; C5 ( Fun/pastime). C6 ( To satisfy ego) ; C7 (Girl students) ; C8 ( Breach of trust. ). Figure 2.2 represents the corresponding adjacency matrix. Given the starting state vector S1 = (0? 0 0? 0 0 1), Inference of the N.C.M of the figure 2.1 gives the scenario depicted in figure ? 0? ? ? 0 1? N(E) =? 0? ? 0 0? Figure 2.2 The adjacency matrix of the Fig. 2.1 We reached a typical behavior, after five iterations, which can be interpreted as follows: S1*N(E) = (0? 0 0? 0 0 1) ( 0? 0 0? 0 1 1) = S2. Iteration 1 S2*N(E) = ( 0?? 0? 0? 1) ( 0?? 0? 0 1 1) = S3. Iteration 2 S3*N(E) = ( 0?? 0??? 1+?) ( 0?? 0?? 1 1) = S4... Iteration 3 S4*N(E) = (??? 0????+1) (??? 0?? 1 1) = S5.Iteration 4 S5*N(E) = (??? 0??? 1) (??? 0?? 1 1) = S6 = S5.Iteration 5.. End Figure 2.3: A Scenario of reasoning on the cognitive map of Figure 2.1 C7 C5 C4 If the "Hackers are girl students then they act of a breach of trust and an indeterminate way of: c1, c2; c3, c5 and c6 and not for a cause of c4. C6 Figure 2.1: Hacking of s by students 4. A case study: The viral infection According to [10], the infection is "Invasion of an alive organism by micro pathogenic organisms (bacteria, virus, mushrooms, parasites). During an infection the pathogenic

144 123 micro-organisms act while multiplying (virulence) and possibly by secreting toxins ". A virus according to [6] is defined by: "Microscopic, simple infectious agent that edge multiply only in living room cells of animals, seedlings, gold bacteria." The process of the infection passes by the seven (7) following stages [6]. a) Attachment of the virus on the surface of the host cell. b) Penetration of the virus and injection of its genetic inheritance inside the cell. c) Transcription of the viral genome to produce of A.R.N- messenger (ARNm) d) Translation of the viral RNAm to proteins. e) Reproduction of the viral genome to form virus genomes. f) Assembly of the genomes wire with proteins produced to give rise to viruses. g) Left the offspring of the virus of the host cell, and the cycle begins again to attack another cell of the same organism or another organism. Two dominating factors play their roles in the viral infection which are the pathogenic force of the virus and the immunizing force of the host cell which determines the infection or the non infection of the latter, as explained in [5 ]: "an infection develops when natural defenses of the organization cannot prevent some; it is the relationship between the quality of the immunizing defenses, more or less compromised during a variable time, and the pathogenic capacity, more or less marked, germ which determines the appearance or not of the infectious disease. Another significant factor which plays its role in the mechanism of the viral infection is well the environment. Indeed much viral infections are of latent type, i.e. without harmful effects on the life of the cell and the organization structurally identical to the parent virus. The actions of the virus depend both on its destructive tendencies toward a specific host cell and on environmental conditions. [6]. Viral infection does not always result in cell death or tissue injury; in fact, most viruses lie dormant in tissue without ever causing pathological effects, or they do so only under other, often environmental, provocations. [6]. Our work is an attempt aiming at bringing into play not only the internal factors (cell-virus) but also the external factors related to the environment in order to better understand this mechanism of the viral infection. All these factors and the relations which can exist among them are modeled graphically by a neutrosophic cognitive map on which will be carried out a causal reasoning making it possible to deduce all its possible behaviors. 5. Implementation We developed a tool for construction and manipulation of Fuzzy and Neutrosophic Cognitive Maps allowing us to reason on complex systems. By way of application, we modeled and simulated the process of the viral infection. With the assistance of experts, we enumerated twenty three (23) different concepts and 30 relations of influence between these concepts describing the viral infection. We give hereafter these concepts as well as the relations of influence between them. C1. Radioactivity C2. Smoking C3. Socio-economic conditions C4. Malnutrition C5. Drug immunosuppressor C6. Congenital disorder of immunity C7. Chemical products C8. Cancer C9. Hygiene C10. Immunodepression C11. Viral infection 12. Bacterial infection C13. Temperature of the body C14. Diabetes C15. Contagious persons C16. Demolition of suspect animals C17. Animal tanks C18. Industry C19. Globalization C20. PH of the virus medium. C21. Vaccination C22. Existence of positive ions around the cell C23. Effet of Greenhouse The qualitative relations between these concepts are given as follows: R1. Radioactivity (+) Cancer R2. Smoking (+) Cancer R3. Chemical products (+) Cancer R4. Cancer (+) Temperature of the body R5. Temperature of the body (+) viral Infection R6. Bacterial infection (+) Temperature of the body R7. Demolition of suspect animals (-) animal Tanks R8. Animal tanks (+) viral Infection R9. Animal tanks (+) bacterial Infection R10. Industry (+) Chemical products R11. Industry (+) the effect of greenhouse R12. Globalization (+) viral Infection R13. Hygiene (-) Smoking R14. Hygiene (-) viral Infection R15. Hygiene (-) Contagious persons R16. PH of the medium of virus (+) viral Infection R17. Vaccination (-) viral Infection R18. Existence of positive ions around the cell (+) viral Infection. R19. Viral infection (+) Temperature of the body R20. The effect of greenhouse (+) viral Infection

145 124 Figure 3: The cognitive map «Viral infection» created by the developed tool. R21. Viral infection (+) Contagious persons R22. Contagious persons (+) viral Infection R23. Socio-economic conditions (-) viral Infection R24. Socio-economic Conditions (-) Malnutrition R25. Socio-economic conditions (+) Hygiene R26. Malnutrition (+) Immunodepression R27. Drug immunosuppressor (+) Immunodepression R28. Congenital disorder of immunity (+) Immunodepression R29. Diabetes (+) Immunodépression R30. Immunodepression (+) Viral Infection. Now, we can build the cognitive map using the developed tool, as shown in figure 3. Using the same tool, we can also obtain all the possible scenarios, also called typical behaviors. These last constitute a subset of the research space consisting of possible cases to consider, i.e more than 8 million cases. As an indication, the time necessary to explore all these cases is 1 hour 20 mn using a computer equipped with a Core2Duo 2GHz processor and a 2Go of RAM. As an example, when greenhouse effect concept is activated, the inference of the cognitive map of the figure 4 (which shows the cognitive map of the virus infection process plus an indeterminate relation between the effect of greenhouse and diabetes) gives the following typical behavior shown in Table 2. This scenario could be interpreted as follows: If effect of greenhouse concept is positively activated then: a) The concepts Viral infection, Temperature of the body and Contagious persons are positively activated. b) The concepts Cancer, Immune- depression and Diabetes are activated with an indeterminate manner. c) The remaining concepts are not activated. 6. Conclusion and future work We developed a tool for reasoning on causal qualitative knowledge represented by neutrosophic cognitive maps. These last are an extension of fuzzy cognitive maps which make it possible to take into account the indeterminism in the study of the complex systems. Like first application, we studied the process of the viral infection. With this intention, we took into account not only factors directly dependent on the cell and the virus but also those factors relating to the environment (such as effect of greenhouse) and the socio economic reality (such as malnutrition and globalization). The tool we developed allows handling and analyzing fuzzy and neutrosophic cognitive maps and it is independent of the field of expertise. Figure 4: Neutrosophic cognitive map (indeterminate relation between «Effect of greenhouse» and «Diabetes») Concept Radioactivity (0) Chemical products (0) Bacterial infection (0) demolition of Suspect animals (0) Smoking (0) Cancer (?) Temperature of the body (+) animal Tanks (0) Globalization (0) Industry (0) Viral infection (+) Vaccination (0) PH of the medium of virus (0) Hygiene (0) Socio- economic conditions (0) Malnutrition (0) Drug immunosuppressor (0) Congenital disorder of immunity (0) Immune depression (?) Diabetes (?) Contagious persons (+) Existence of pos. ions around cell (0) greenhouse effect (+) Effect no effect no effect no effect no effect no effect indeterminate positive no effect no effect no effect positive no effect no effect no effect no effect no effect no effect no effect indeterminate indeterminate positive no effect positive Table 2: An execution scenario of the developed viral infection system.

146 125 As a future work, we plan to study the following items: a) It was noticed that the inference on the cognitive maps consists of finding an answer to this question "what will happen if?", but in practice the decision maker wants often to know an answer to the question "if I want to, what should I make? ". For example in our case "the viral infection", one noticed that the decision maker puts this question intuitively " If I want to decrease the viral infection, what have I to do? With this intention, an algorithm for the abductive causal reasoning is to be proposed. b) The space of research of the typical behaviors in a cognitive map increases exponentially with the number of concepts considered. It would be then very interesting to find heuristics in the aim to attenuate this problem of complexity. c) Modeling and simulating a distributed system using neutrosophic cognitive map model (like a weather prediction system) by means of a multi agent system. This work will constitute a very good example of a collective and qualitative decision making. [11] Elpiniki I. Papageorgiou, Nikolaos I. Papandrianos, Georgia Karagianni, George C. Kyriazopoulos and Dimitrios Sfyras. A Fuzzy Cognitive Map based tool for prediction of infectious diseases, Inter. Journal of expert systems with Applications, vol. 36, issue 10, pp , Tahar Guerram is an Assistant Professor of Computer Science at the Department of Computer Science of the University of Oum El- Bouaghi in Algeria. He is a Phd s degree student at University of Constantine in Algeria. His main areas of interest include agent based knowledge discovery and qualitative reasoning in multi agent systems. Ramdane Maamri is an Assistant Professor of Computer Science at the Department of Computer Science of the University Mentouri of Constantine in Algeria. He holds a Ph.D. in Computer Science from the University of Mentouri of Constantine in Algeria. His main areas of interest include Artificial Intelligence and Agent-oriented Software Engineering, Zaidi Sahnoun is a Professor of Computer Science at the Department of Computer Science of the University of Mentouri of Constantine in Algeria. He holds a Ph.D. in information technology from Rensselaer Polytechnic Institute in U.S.A. His main areas of interest include software engineering, formal methods, multi agent sytems and complex systems. References [1] W. B. Vasantha Kandasamy and Florentin Smarandache. Fuzzy Cognitive Maps and Neutrosophic Cognitive Maps, Xiquan, 510 E. Townley Ave., Phoenix, AZ 85020, USA; 2003 [2] Axelrod, R. Structure of Decision. Princeton, New Jersey, Princeton University Press [3] Kosko, B. Fuzzy Cognitive Maps, International. journal. of Man-Machine Studies, vol. 24, pp , [4] Stylios, C.D., and P.P. Groumpos. The Challenge of Modeling Supervisory Systems using Fuzzy Cognitive Maps, J. of Intelligent Manufacturing, vol. 9,pp ,1998. [5] Petit Larousse de la Médecine, printed in France, Juin 2003, ISBN : [6] Encyclopedia Britannica online, website: [7] E.I. Papageorgiou, C.D. Stylios, P.P. Groumpos, An integrated two-level hierarchical decision making system based on fuzzy cognitive maps, IEEE Transactions on Biomedical Engineering, vol. 50, pp , [8] V.C. Georgopoulos, G.A. Malandraki, C.D. Stylios, A fuzzy cognitive map approach to differential diagnosis of specific language impairment, Journal of Artificial Intelligence in Medicine, vol. 29, pp , 2003 [9 ] E.I. Papageorgiou, P. Spyridonos, C.D. Stylios, R. Ravazoula, P.P. Groumpos, G. Nikiforidis, A soft computing method for tumour grading cognitive maps, Journal of Artificial Intelligence in Medicine, vol. 36, pp , [10] Hailin Feng and Chenxi Shao, Using Qualitative Technology for Modeling the Process of Virus Infection. Life Systems Modeling and Simulation, LNCS, vol. 4689, pp , 2007.

147 ISSN (Online): Health Smart Home Ahmad CHOUKEIR 1, Batoul FNEISH 1, Nour ZAAROUR 2, Walid FAHS 1, Mohammad AYACHE 2 1 Department of Computer and Communication, Islamic University of Lebanon, Faculty of Engineering Khaldeh, Lebanon 2 Department of Biomedical, Islamic University of Lebanon Faculty of Engineering Khaldeh, Lebanon Abstract We present in this paper a study and an experimental medical tele-surveillance system for maintaining patients at home. The aim of this paper is to demonstrate how combining many kinds of technologies starting with sensors connected to the patient then using wireless technology (ZigBee) to transmit information and finally analysis and detection of critical situation such as falls, ECG problems, hypothermia hyperthermia by using a GUI created to fulfill these missions. Many physiological parameters can be studied but we tended in our project to study ECG signal, heart rate, temperature of the patient, fall detection and medication reminder. Results obtained allow the user to visualize online the parameters and to detect any problem that occurs, so by using this system we can observe the status of the patient during his normal life at his home. personal reasons because the ability of the smart home to measure medical data, and send it to the medical center through a network, makes it possible to lengthen the independent living at home. The diagram described below aims to show how Health Smart Home contributes to increase the quality of life of patients and how telemedicine allows helping diagnostics of diseases without retardation to detect the minimal deterioration in patient s status. Keywords Sensors, network, Telemedicine, Smart Home, VB.net. 1. Introduction The term smart home [1, 2] has been used for more than a decade to introduce the concept of networking devices and equipment in the house. The demographic trends of the advanced countries go towards the development of an older population so Health Smart Homes represent one of the most promising ways of development of telemedicine. The aim of Smart Homes is the earlier detection and treatment of diseases. This need is due to many reasons such as economical reasons because earlier detection and treatment of diseases can be the solution to decrease the high costs, medical reasons (the most important) to monitor remotely the status of the patient living alone to detect possible deterioration in his health status, and Fig. 1: Bloc diagram 2. Materials and Methods 2.1. The chain of information and communication 2.2. This paper shows different steps followed to achieve this project. The underlying model of the whole human devices system addresses concepts on 5 different levels as follows:

148 ISSN (Online): Sensors level (Level 1): the perception level of physical events by means of sensors. - Transmission level (Level 2): the level of information s distribution. - Software level (Level 3): the level of communication software used in the project. - Network level (Level 4): the level of sharing data concerning the system between the Smart Home and the Medical Center. - Human level (Level 5): we define at that level a series of typical profiles of people involved in the medical tele-observation Description of levels Level 1: Sensors level Many physiological parameters can be studied to observe the status of the patient circuit LM35 with an amplifier and a simple calculation is made to know the temperature proportional to the voltage at the output of the circuit. Equilibrium and fall in case of an elderly person which can result in quite serious injuries because of his/her fragile bone structure. In-fact the detection of the fall as soon as possible can help to limit the negative consequences. To detect the fall of the patient we used a key fill at a determined level with mercury which is a metal. This key is placed on the hip of the patient in case if he falls the metal will close the circuit and a voltage of 5 volt will be obtained at the output of the circuit. In addition to, the medication is designed to improve patient health, but it must be taken correctly and in time to be effective. So Health Smart Home is charged to remember the patient by his medication including all details (name, time, number of pills and duration). Electrocardiogram (ECG) which represents measures and records the electrical activity of the heart in exquisite detail. Interpretation of these details allows diagnosis of a wide range of heart conditions and computing the Heart Rate which is the number of heartbeats per unit of time - typically expressed as beats per minute (bpm) - which can vary as the body's need for oxygen changes, such as during exercise or sleep. The measurement of heart rate is used by medical professionals to assist in the diagnosis and tracking of medical conditions. To achieve the acquisition of the signal ECG from the patient we made an electronic circuit to detect the ECG of the patient.this work is done by the implementation of the circuit below. Level 2: Transmission level To transmit the information from the sensors to the computer at Patient Home we used the modules ZigBee to fulfill this mission. Zigbee is a software-based protocol that sits on top of the RF wireless devices standard similar to Bluetooth. Unlike Bluetooth, Zigbee is capable of forming large networks of nodes and boasts advanced features such as mesh networking, simple addressing structures, route detection, route repair, guaranteed delivery and low power operation modes. [3,4]. In this paper we will describe two circuits used to fulfill this job: 1 Emission circuit Receiving circuit 1. Emission circuit This circuit is in the pocket carried by the patient and its mission is to transmit the data to the receiving circuit. The module ZigBee should be configured as an emitter. A microcontroller is used to organize the data whether it refers to an ECG signal, temperature or fall. Fig.2: ECG circuit Temperature which represents an important parameter to determine the case of the patient whether it s a case of hypothermia, hyperthermia or normal case. This physiological parameter is measured by using the integrated

149 ISSN (Online): Fig. 3: Emission circuit 2. Receiving circuit This circuit is related to the computer of the patient to transmit serially the data. The module ZigBee should be configured as a receiver. Fig.4 Receiving circuit parameters ( ECG signal, heart rate, temperature of the patient and its description, the status of the patient (in case of fall or not)). In addition to, the application allows alerting the medical staff in case of problems or deterioration of the patient s state. Data base The application includes many data bases concerning all the information related to the patient and to the system (problems, registration, updating..). The figure fig. 5 shows the data bases used by the GUI. The module A described fills the data base concerning the problems. Besides the module B fills the data bases concerning the patient and the medical staff, also this module uses the data bases to assure the security of the system and tells which computers on the system are using the application and the computers which can be alerted in case of problems. Level 3: Software level Description of the application In this level an application is created using Vb.net to accomplish many missions: 1. To get the information from the sensors (physiological parameters). 2. To enter all data concerning a patient or an employee. 3. To ensure system security. 4. To give a level of access for each user. 5. To identify problems. This application includes two types of modules. Module A: This module is located at the patient's home and devoted to information acquisition, data recording, analysis, testing and filling a data base in case of problems. This information is shared on the network. Module B: this module is located at the patient s home and at the medical center. At the patient s Home: the aim of the application in this case is to remember the patient by his medication and to give him the ability to visualize all details about his medications. At the medical center: The main functions available through the GUI according to the user's access rights are: allowing the registration of new patients or members in medical staff to give them the authorization to have an access on the system with respect to their privilege, allowing to the doctors the prescription of the medications for a specific patient, write/display a report or a prescription, monitor sensors and display statistics allowing the visualization of all the Level 4: Network Fig. 5: Data Base In order to transmit the information from the patient s computer to the medical center and vice versa a network is needed to accomplish this mission so the files undergo the principle File sharing which is the act of making specified data files accessible to others. File Sharing can be done publicly such as on the Internet, or privately within a network. The files to be shared can come from a pc or a server. Access can be controlled and vary amongst clients, and files can be restricted from editing and a map network drive is used to transform the files to a new driver to simplify the work.

150 ISSN (Online): doesn t belong to the normal range an alert will be activated in the medical center. (Figure 7) Fig. 6: Sharing files and Mapping Level 5 : Human level The access to the system is allowed to authorized users only so in our project we define four user profiles: administrator, medical, partial and patient. One physical person may have one or several user profiles. Table 1: Possible set up of access rights Fig. 7: ECG information The temperature is measured and tested if it belongs to the normal range and the result is visualized (figure 8). Fig.8 Temperature Information Each access determines the behavior of the user s application and that is used to assure a high level of security and confidentiality of the system. 3. Results The first result to mention is the realization of an operational information and Communication System-Health Smart Home demonstration system. We realize all the electronic circuits described above. Also a GUI is created using VB.net with 3136 program lines and 45 functions. We could observe the electrocardiogram online and calculate the heart rate each 60 seconds, in case if the heart rate The fall of the patient is detected and a camera is turned on and a figure is taken (figure 9). Fig.9 Fall Info

151 ISSN (Online): On the other hand forms are created in the application to accomplish the mission of update system, registration, networking, prescription of the medication, and searching for the coordinates of an employee or a patient. The most important mission achieved by the application is alerting the medical staff in case of any problem and this is accomplished by a Problem Form which determines all the details of the problem. 4. Conclusion and future work The real use of this system in the everyday life of a person meets a certain number of issues at the following general levels: moral, ethical and medico legal, socio-medical and organizational, economical, technical and of quality assurance. [5, 6, 7]. It was our first step in telemedicine and all this work has been done in a short time (four months including researches, planning, programming, and implementation and getting results in real time). This was a simple base which can be developed by using many sensors to reach an Ubiquitous acquisition of the data to observe the detailed status of the patient. Acknowledgment We are grateful to the direction of the faculty of Engineering at the Islamic University in Lebanon. This project is done in the laboratory of the faculty. References [1] Brooks RA. The Intelligent Room Project. In: Marsh JP, Nehaniv CL, Gorayska B, editors. Proc. CT'98 : 2nd Int Conf Cognitive Technology. Los Alamitos, California: IEEE Computer Soc; pp [2] Tang P. and T. Venables, 'Smart' homes and telecare for independent living, J Telemed Telecare2000: 6(1) pp [3] Drew Gislason, Zigbee Wireless networking, Elsevier, [4] Ata Elahi, Adam Gschwender, Zigbee Wireless Sensor and control Network, ISBN 13: [5] An Experimental Health Smart Home and Its Distributed Internet-based Information and Communication System: First Steps of a Research Project Vincent Rialle, Norbert Noury, Thierry Hervé Laboratory TIMC-IMAG UMR CNRS 5525, μisfv team, Grenoble, France [6] Fisk MJ. Telecare equipment in the home. Issues of intrusiveness and control. J Telemed Telecare 1997: 3(Suppl 1) pp [7] McShane R., T. Hope and J. Wilkinson, Tracking patients who wander: ethics and technology, Lancet 1994: 343 pp. 1274

152 SD-AODV: A Protocol for Secure and Dynamic Data Dissemination in Mobile Ad Hoc Network 131 Rajender Nath 1 and Pankaj Kumar Sehgal 2 1 Associate Professor, Department of Computer Science and Applications, Kurukshetra University, Kurukshetra, Haryana, India 2 Assistant Professor, Department of Information Technology, MM Engineering College, MM University, Ambala, Haryana, India Abstract Security remains as a major concern in the mobile ad hoc networks. This paper presents a new protocol SD-AODV, which is an extension of the exiting protocol AODV. The proposed protocol is made secure and dynamic against three main types of routing attacks- wormhole attack, byzantine attack and blackhole attack. SD-AODV protocol was evaluated through simulation experiments done on Glomosim and performance of the network was measured in terms of packet delivery fraction, average end-to-end delay, global throughput and route errors of a mobile ad hoc network where a defined percentage of nodes behave maliciously. Experimentally it was found that the performance of the network did not degrade in the presence of the above said attacks indicating that the proposed protocol was secure against these attacks. Keywords: Nework Security, Routing Attacks, Routing Protocol, Simulation Experiments. 1. Introduction A multi-hop mobile ad hoc network (MANET) consists of a group of mobile wireless nodes that self configure to operate without infrastructure support. Network peers communicate beyond their individual transmission ranges by routing packets through intermediate nodes. Security remains as a concern in MANET. In general, a MANET is vulnerable due to its fundamental cooperation of open medium, absence of central authorities, dynamic topology, distributed cooperation and constrained capability [1]. A node in the MANET without any adequate protection can become an easy target for attacks. Attacker just needs to be within radio range of a node in order to intercept the network traffic. The attacks on MANET are classified as passive attacks and active attacks [22]. In passive attacks, an intruder snoops the data exchanged between the nodes without altering it. In these type of attacks, a selfish node abuses constrained resources such as battery power for its own benefit. The goal of an attacker is to obtain the information that is being transmitted that leads to the violation of massage confidentiality. Passive attacks are difficult to detect because the activity of the network is not disrupted in these attacks. In active attacks, an attacker actively participates in disrupting the normal operation of the network services. These can be performed by injecting incorrect routing information to poison the routing table or by creating a loop. These attacks are further divided into external and internal attacks. External attacks are carried by nodes that are not authorized part of the network. Internal attacks come from compromised nodes, which are legitimate part of the network. Active attacks are very difficult to detect because the attacker is part of the network. There are basically two approaches to securing a MANET: proactive and reactive. The proactive approach attempts to prevent security attack, typically through various cryptographic techniques. On the other hand, the reactive approach finds an attack and reacts accordingly. Both approaches has there own merits and suitable for different issues of security in MANET. Most of the secure routing protocols adopt proactive approach to securing routing control messages and reactive approach to secure data packet forwarding messages. A complete security solution requires both proactive and reactive approaches. While a number of routing protocols [3-11] have been proposed by the Internet Engineering Task Force s MANET working group but they are silent in terms of security. Most of the MANET secure routing protocols have been proposed in the literature such as SEAD [12], ARIADNE [13], SAR [14], SRP [15], CONFIDANT [16], ENDAIRA [17], TESLA [21] etc. do not mitigate against

153 132 these attacks. Some solutions against particular attacks have been presented by the researchers such as rushing attack and defenses [18], wormhole attack and defenses [19], sybil attack and defenses [20]. Because these solutions are designed explicitly with certain attack models in mind so they work well in the presence of designated attacks but may collapse under unanticipated attacks. Therefore, a more ambitious goal for MANET security is to develop a multifence security solution that can offer multiple lines of defenses against both known and unknown security threats. Rest of the paper is structured as follows: Section 2 discusses the base routing protocol AODV. Section 3 describes the new protocol SD-AODV. Section 4 describes and compares the simulation experiment and result performed on AODV and SD-AODV protocol in presence of malicious nodes. Section 5 gives the concluding remarks. 2. Ad hoc On-demand Distance Vector (AODV) Routing Protocol AODV is an improvement on DSDV [23] because it typically minimizes the number of required broadcasts by creating routes on a demand basis. AODV routing protocol uses reactive approach for finding routes, that is, a route is established only when it is required by any source node to transmit data packets. The protocol uses destination sequence numbers to identify the recent path. In this protocol, source node and the intermediate nodes store the next node information corresponding to each data packet transmission. In an on-demand routing protocol, the source node floods the Route REQuest (RREQ) packet in the network when a route is not available for the desired destination. It may obtain multiple routes to different destinations from a single RREQ. A node updates its path information only if the destination sequence number of the current packet received is greater than the last destination sequence number stored at the node. are discarded. All intermediate nodes having valid routes to the destination, or the destination node itself are allowed to send RREP packets to the source. Every intermediate node, while forwarding a RREQ, enters the previous node address and its BcastID. A timer is used to delete this entry in case a RREP is not received before the timer expires. This helps in storing an active path at the intermediate node as AODV does not employ source routing of the data packets. When a node receives a RREP packet, information about the previous node from which the packet was received is also stored in order to forward the data packet to this next node as the next hop towards the destination. 3. Proposed SD-AODV Protocol The existing AODV protocol is not secure against any routing attack. We have extended the existing AODV protocol to make it secure against the three types of routing attacks- Wormhole attack, Byzantine attack and Blackhole attack. The proposed protocol is named SD- AODV (Secure and Dynamic Ad Hoc On-Demand Distance Vector) which is secure and dynamic against In following paragraphs we describe different schemes to make the protocol secure against the above said three attacks. Let N= {n 1, n 2, n 3,.,n k } is a set of k nodes in the network that includes destination nodes and malicious node. Let D={d 1,d 2,d 3,.,d j } is a set j destination nodes where D N, j<k and M={m 1, m 2, m 3,,m h }is set of h malicious nodes where M N, h<k. Any member of M can act as a malicious node to perform either wormhole attack malicious node or byzantine malicious node or blackhole malicious node. The three different schemes has been formed and tested to safeguard against three different attacks. The packet has been transferred to Node G instead of Node H A RREQ carries the source identifier (SrcID), the destination identifier(destid), the source sequence number (SrcSeqNum) and destination sequence number (DestSeqNum), the broadcast identifier (BcastID), and the time to live (TTL) field. DestSeqNum shows the freshness of the route that is selected by the source node. When an intermediate node receives a RREQ, it either forwards it or prepares a route reply (RREP) if it has a valid route to the destination. The validity of a route at the intermediate node is determined by comparing the sequence number at packet. If a RREQ is received multiple times, which is indicated by BcastID-SrcID pair, then the duplicate copies c B D A E F Figure 1. Wormhole Attack G I H

154 133 The first scheme makes the proposed protocol secure against the wormhole attack. In the wormhole attack, an attacker receives packets at one location in the network and tunnels them to another location in the network. The tunnel between these two locations is referred as a wormhole. Due to the broadcast nature of the radio channel, the attacker can create a wormhole even for packets not addressed to itself. In Figure 1, let us assume node A is source node, H is destination node and E is malicious node, which can commit wormhole attack. When node A broadcasts a RREQ packet to its neighbors B, C and E, then the malicious node E commit wormhole attack and changes its destination address to G hence the request does not reach to the destination node H. To safeguard against the wormhole attack we make use of hash chains in the proposed protocol. Hash chains are created by applying a secure hash algorithm [24]. This scheme is used for protecting the portion of the information in the RREQ messages, which is destination address. The protocol computes the digest (DigestAddr) of the destination address and appends that with the RREQ packet. It computes and compares the destination address at every intermediate node. If the destination address has been changed by the intermediate node, then it declares the node as malicious node and change the destination address to its original. Steps of the scheme are summarized below: Step 1: Compute DigestAddr of destination address by calling SHAReset ( ) function and SHAInput ( ) function Step 2: Append computed digest in DigestAddr of Data packet. Step 3: if (n x is an Intermediate Node) then Fetch Destination address from Data packet and compute digest named New_computed_Digest by calling Call SHAReset ( ) function and Call SHAInput ( ) function Step 4: Compare DigestAddr with new computed digest in Step 3 if(digestaddr=new_computed_digest) then o m z node is detected as malicious node o SD-AODV change destination address field to accurate destination address The second scheme makes the proposed protocol secure against the byzantine attack. In the byzantine attack, a malicious node or a set of malicious node works in collusion and carries out attacks such as creating routing loops and routing packets on non-optimal paths. It consumes energy and bandwidth of the network. In Figure 2, let us again assume node B is source node, H is destination node and C is malicious node, which can commit routing loop attack. When node B wants to transmit a data packet to the destination node H, the malicious node C loops the information back to node A as shown in Figure 2 and the packet does not reach to the destination node H. F A B Next hop C G C Change next hop Figure 2. Byzantine Attack Next hop A SD-AODV protects the mutable information such as Nexthop field in the RREQ. This assures that nodes receiving AODV messages that the Next-hop value provided are accurate and have not been changed by the malicious node on the path. Every time when a node route packet to MAC layer the protocol dynamically save the value of Next-hop field. Whenever any intermediate node receives an RREQ message, the protocol verifies the Next-hop values. If the Next-hop value has been changed, the node is declared as a malicious node. The protocol then changes the Next-hop value to its original. Steps of the scheme are summarized below: The third scheme makes the proposed protocol secure against the blackhole attack. In the blackhole attack, a malicious node falsely advertises good paths (e.g. shortest D the The data packet has been looped back by node C to A Step 1: Everytime when RoutePacketAndSendToMac( ) called SD-AODV save the value of Next-hop dynamically. Step 2: if (n x is an Intermediate Node and d y exist in route table) then If (Address of Next-hop has been altered) then m z node is detected as malicious node SD-AODV changes its value to original by updating route table. H I E

155 134 path or most established path) to destination node during the route request phase. The intention of malicious node could be to hinder the route request phase or to stop all data packets being sent to the actual destination node. In Figure 3, let us again assume node A is source node, D is destination node and G is malicious node, which can commit routing blackhole attack. When node A wants to transmit a route request packet to the destination node H, the malicious node E advertise itself as shortest path and sends route reply shown in Figure 3 and the packet does not reach to the destination node H. RREQ A RREP B RREQ G The node G advertises itself as shortest path and sends route reply C Figure3. Blackhole Attack SD-AODV dynamically protects the route establishment process, by cross verifying the RREP (Route reply) message to ensure that the sender will not receive any RREP from malicious node. Steps of the scheme are summarized below: Step 1: if (n x is a Intermediate Node) then Call RoutingAodvCheckRouteExist( ) if (d y does not exists in Route Cache) then m z node is detected as malicious node Step 2: SD-AODV stop the route establishment process by calling RoutingAodvInitiateRREPbyIN ( ) RoutingAodvRelayRREQ(..) 4. Simulation Experiment and Results A simulation experiment was performed by using Glomosim [2] simulator to study the effects of all the three attacks mentioned above on the proposed protocol: SD- AODV. The simulation experiment was performed on a computer with Intel core 2 Duo 1.7 GHz processor and 2GB RAM. The simulation experiment was performed E D H twice by taking 50 and 100 node to study the effects of the three attacks by measuring the performance of the network. In each of the case i.e. 50 nodes and 100 nodes simulation was carried out several times with different seed values. Other parameters that were taken for simulation are shown in the table 1. Table 1: Simulation Parameter Value Parameter Vale Description Terrain Range 1KM 1KM X,Y Dimension in Kilometer Power Range 250m Nodes s power range in meters Simulation 100 s Simulation duration in Time seconds Node Uniform Node placement policy Placement Mobility 5m/s-20 m/s Random Waypoint in meter per second Traffic Modal CBR Constant bit rate Packet Size 512 Bytes Minimum transfer unit MAC IEEE Medium Access Control Protocol Bandwidth 2MBPS Node s Bandwidth in Mega Routing Protocol AODV bits per second Base routing protocol for ad hoc networks The same experiment was repeated for the existing protocol AODV in order the compare it with the proposed protocol SD-AODV. 4.1 Testing SD-AODV and AODV Protocols against Wormhole, Byzantine and Blackhole attacks The performance of the network was evaluated by using the following four metrics (i) Packet Delivery Fraction (PDF), (ii) Average End-to-End Delay, (iii) Throughput and (iv) Route error. Packet Delivery Fraction (PDF): It is a ratio of the data packets delivered to the destinations to those generated by the Constant Bit Rate (CBR) sources. Average End-to-End Delay: This includes all possible delays caused by buffering during route discovery latency, queuing at the interface queue, retransmission delays at the MAC, and propagation and transfer times. Throughput: It is equal to the average performance of all nodes during simulation. It is a calculation of bits per second processed by each node. Route Errors:The error messages garneted by the protocol during simulation.

156 135 Simulation was performed by taking different seed values. In the experiment, the numbers of malicious node were increased starting from 5% to maximum of 30 % in the step of 5%. The Glomosim simulator generated a GLOMO.STAT file which contained all the statistics regarding number of packets send, number of packets received, number of bytes sents, number of bytes received, throughput(bits per second), delay (in seconds), number of route errors etc Using Packet Delivery Fraction (PDF) PDF was calculated by extracting data from the GLOMO.STAT file and four curves (one for wormhole attack on AODV, one for byzantine attack on AODV and one for blackhole attack on AODV and one for three attack on SD-AODV) are plotted by taking %age of malicious node on X-axis and %age of PDF on Y-axis as shown in Figure 4(a) and 4(b) for 50 nodes and 100 nodes respectively. From the Figure 4(a) and (b), it is quite clear that SD-AODV not only prevents from various attacks but also gives better performance while increasing the number of malicious nodes. the average delay has increased in the case of SD-AODV due to overhead increased for protection against three attacks. But it is still less as compare to Byzantine attack on AODV because it loops back data and utilize resources and bandwidth %age PDF Wormhole Attack on AODV Byzantine Attack on AODV Blackhole Attack on AODV SD-AODV %age Malicious Node Figure 4(b). PDF Comparison of different attacks on AODV and SD-AODV in 100 nodes scenario %age PDF Average delay(sec.) Wormhole Attack on AODV Byzantine Attack on AODV Blackhole Attack on AODV SD-AODV Wormhole Attack on AODV Byzantine Attack on AODV Blackhole Attack on AODV SD-AODV %age Malicious Node Figure 4(a). PDF Comparison of different attacks on AODV and SD-AODV in 50 nodes scenario Average End-to-End Delay Average end-to-end delay was calculated by extracting data from the GLOMO.STAT file and four curves (one for wormhole attack on AODV, one for byzantine attack on AODV and one for blackhole attack on AODV and one for three attack on SD-AODV) are plotted by taking %age of malicious node on X-axis and Average delay on Y-axis as shown in Figure 5(a) and 5( b) for 50 nodes and 100 nodes respectively. The average delay has been increased almost double in case of 100 nodes as compared to 50 nodes. From the Figure 5(a) and 5(b), it is quite clear that %age Malicious Node Figure 5(a). Average End-to-End Delay Comparison of different attacks on AODV and SD-AODV in 50 nodes scenario Throughput Throughput was calculated by extracting data from the GLOMO.STAT file and Four curves (one for wormhole attack on AODV, one for byzantine attack on AODV and one for blackhole attack on AODV and one for three attack on SD-AODV) are plotted by taking %age of malicious node on X-axis and throughput (bits per second) on Y-axis as shown in Figure 6(a) and 6(b) for 50 nodes and 100 nodes respectively. It is obvious from the figures that, throughput is increased in case of SD-

157 136 AODV due to more calculation work for protective measures. Average delay(sec.) Wormhole Attack on AODV Byzantine Attack on AODV Blackhole Attack on AODV SD-AODV %age Malicious Node Figure 5(b). Average End-to-End Delay Comparison of different attacks on AODV and SD-AODV in 100 nodes scenario Route Errors Number of route errors was calculated by extracting data from the GLOMO.STAT file and Four curves (one for wormhole attack on AODV, one for byzantine attack on AODV and one for blackhole attack on AODV and one for three attack on SD-AODV) are plotted by taking %age of malicious node on X-axis and number of route errors on Y-axis as shown in Figure 6(a) and 6(b) for 50 nodes and 100 nodes respectively. It is quite clear that number of route errors have drastically increased in the case of blackhole attack because it attacks on route establishment process. Route errors are nominal in case of SD-AODV Throughput (bits per second) Wormhole Attack on AODV 3600 Byzantine Attack on AODV Blackhole Attack on AODV SD-AODV %age Malicious Node Figure 6(a). Throughput Comparison of different attacks on AODV and SD-AODV in 50 nodes scenario No. of route errors Wormhole Attack on AODV Byzantine Attack on AODV Blackhole Attack on AODV SD-AODV %age Malicious Node Figure 7(a). No. of Route Error Comparison of different attacks on AODV and SD-AODV in 50 nodes scenario Throughput (bits per second) Wormhole Attack on AODV Byzantine Attack on AODV Blackhole Attack on AODV SD-AODV No. of route errors Wormhole Attack on AODV Byzantine Attack on AODV Blackhole Attack on AODV SD-AODV %age Malicious Node %age Malicious Node Figure 6(b). Throughput Comparison of different attacks on AODV and SD-AODV in 100 nodes scenario Figure 7(b). No. of Route Error Comparison of different attacks on AODV and SD-AODV in 100 nodes scenario

158 137 References 4. Conclusion In this paper, we have presented a protocol which is secure & dynamic against three types of attacks - wormhole attack, Byzantine attack and blackhole attack. The protocol is based on the existing protocol AODV. The proposed protocol has been tested on Glomosim by using four metrics- packet delivery Fraction, average end-to-end delay of data packets, throughput and route error against widely used protocol AODV. The packet delivery fraction (PDF) metric has shown that all the three routing attacks sharply decrease the PDF performance in the case of AODV protocol but in case of the proposed protocol SD-AODV there is no fall in the PDF, which has clearly indicated that the proposed protocol has became secure against the three attacks in question. The Average end-to-end delay metric has shown that average delay has increased in the case of proposed protocol when the node has got the blackhole attack and wormhole attack, but it is lesser in case of Byzantine attack. The straight increase in delay can be attributed to the overhead incurred due the implementation of additional functionality of the SD-AODV protocol. The result of throughput metric has shown that it is higher in case of the SD-AODV protocol as comparison to AODV protocol in the presence of the three attacks. This again indicates that the proposed protocol has become secure against these attacks. The route errors have drastically decreased in case of SD- AODV protocol, but the results have also shown that the route errors are slightly more in case of SD-AODV as compared to AODV protocol when wormhole attack takes place because in wormhole attack the destination address is changed cleanly without affecting the route. In nutshell, we can say that the proposed SD-AODV protocol has become secure and dynamic against these three attacks. [1] Farooq Anjum and Petros Mouchtaris, Security For Wireless ad Hoc Networks, Wiley interscience. John Wiley & Sons, INC., publication. [2] X. Zeng, R. Bagrodia and M. Gerla, Glomosim: A library for parallel simulation of large-scale wireless networks, in: Proceedings of the 12th Workshop on Parallel and Distributed Simulations PADS 98 (May 26 29, Alberta, Canada, 1998). [3] C.E.Perkins and E.M.Royer, Ad-hoc On-Demand Distance Vector Routing, Proc. 2nd IEEE Workshop of Mobile Comp. Sys. and Apps., Feb. 1999, pages [4] J. Macker et Al., Simplified Multicast Forwarding for MANET, draft-ietf-manet-smf-07, February 25, [5] I. Chakeres et. Al., Dynamic MANET On-demand (DYMO)Routing,draft-ietf-manet-dymo-14, June 25, [6] I. Chakeres et. Al., IANA Allocations for MANET Protocols,draft-ietf-manet-iana-07, November 18, [7] T. Clausen et. Al., The Optimized Link State Routing Protocol version 2, draft-ietf-manet-olsrv2-06, June 6, [8] T. Clausen et. Al., Generalized MANET Packet/Message Format, draft-ietf-manet-packetbb-13, June 24, [9] T. Clausen et Al., Representing multi-value time in MANETs, draft-ietf-manet-timetlv-04, November 16, [10] T. Clausen et Al., MANET Neighborhood Discovery Protocol (NHDP), draft-ietf-manet-nhdp-06, March 10, [11] D. Johnson and D. Maltz., Dynamic source routing in adhoc wireless networks routing protocols, In Mobile Computing, pages 153{181. Kluwer Academic Publishers, [12] Y. C. Hu, D. B. Johnson and A. Perrig, SEAD: Secure Efficient Distance Vector Routing for Mobile Wireless Ad Hoc Networks, Fourth IEEE Workshop on Mobile Computing Systems and Applications (WMCSA 02), Jun [13] Y. C. Hu, D. B. Johnson, and A. Perrig, Ariadne: A Secure On-Demand Routing Protocol for Ad Hoc Networks, Mobicom 02, [14] R. Kravets, S. Yi, and P. Naldurg, A Security-Aware Routing Protocol for Wireless Ad Hoc Networks, In ACM Symp. on Mobile Ad Hoc Networking and Computing, [15] P. Papadimitratos and Z. J. Haas, Secure Routing for Mobile Ad hoc Networks, In Proc. of the SCS Communication Networks and Distributed Systems Modeling and Simulation Conference (CNDS 2002), Jan [16] S. Buchegger and J. L. Boudec, Performance Analysis of the CONFIDANT Protocol Cooperation Of Nodes Fairness In Dynamic Ad-hoc NeTworks, In Proc. Of IEEE/ACM Symposium on Mobile Ad Hoc Net- working and Computing (MobiHOC), Jun [17] Gergely Acs, Levente Buttyan and Istvan Vajda, Provably Secure On-demand Source Routing in Mobile Ad Hoc Networks, IEEE transactions on Mobile Computing, Vol.5, No.11, November 2006, pp [18] Y. -C. Hu, D. B. Johnson, and A. Perrig, Rushing Attacks and Defense in Wireless Ad Hoc Network Routing Protocols, WiSe 2003, [19] Y. -C. Hu, A. Perrig, and D. B. Johnson, Packet Leashes: A Defense against Wormhole Attacks in Wireless Networks, Infocom 2003.

159 138 [20] J. Newsome, E. Shi, D. Song, and A. Perrig, The Sybil Attack in Sensor Networks: Analysis & Defenses, Proc. of the 3rd Intl. Symp. on Information Processing in Sensor Networks, [21] A. Perrig, R. Canetti, D. Tygar, and D. Song, The TESLA Broadcast Authentication Protocol, RSA Cryptobytes (RSA Laboratories), Vol 5, No 2, Summer/Fall 2002, pp [22] C. S. R. Murthy and B. S. Manoj, Ad Hoc Wireless Networks: A Architectures and Protocols, Prentice Hall PTR, [23] C Perkins and P. Bhagwat, Highly dynamic destination sequenced distance vector routing (DSDV) for mobile computers, ACM SIGCOMM, (October 1994). [24] X. Wang, Y. L. Yin, and H. Yu, Finding Collisions in the Full SHA-1, CRYPTO, 2005.

160 139 A Radio Based Intelligent Railway Grade Crossing System to Avoid Collision Sheikh Shanawaz Mostafa 1, Md. Mahbub Hossian 2, Khondker Jahid Reza 3, Gazi Maniur Rashid 4 1 Electronics and Communication Engineering Discipline, Khulna University, Khulna-9208, Bangladesh. 2 Electronics and Communication Engineering Discipline, Khulna University, Khulna-9208, Bangladesh. 3 Electronics and Communication Engineering Discipline, Khulna University, Khulna-9208, Bangladesh. 4 Electronics and Communication Engineering Discipline, Khulna University, Khulna-9208, Bangladesh. Abstract Railway grade crossing is become the major headache for the transportation system. This paper describes an intelligent railway crossing control system for multiple tracks that features a controller which receives messages from incoming and outgoing trains by sensors. These messages contain detail information including the direction and identity of a train. Depending on those messages the controller device decides whenever the railroad crossing gate will close or open. Keywords: Grade crossing, intelligent system, messages, multiple tracks, sensors 1. Introduction A railway grade crossing is a point at which a railway and a road intersect on the same level [4].In many situations two or more tracks may cross a highway. It creates problem for a bus or truck or other vehicles to clear both tracks. In recent five years, more than three hundred people are killed and another about two hundred are injured due to railway grade crossing accidents [5]. Whether accidents are caused by the negligence of the men, undesirable weather conditions, inadequate traffic planning. A large number of grade crossings exist without any gate. There is an inherent unreliability in the present manual system. Other hand, constant warning time system provides fixed warning time regardless of an approaching train. So that, for a faster train the safety devices are activated earlier and for a slower train activated later. Installing and maintaining costs are too much for these above two systems. Global Positioning System (GPS) based rail crossing system is discussed by many authors. Some author discusses the use of data obtained from GPS devices located on trains or at railroad crossings to provide train s approaching information [2]. Other installed GPS receiver is on top of a train and used to obtain information concerning the train s speed and position [3]. And a radio link based system describe in by [1].But all the systems are not useful for deriving arrival time and train speed information for multiple trains at a time. And they cannot derive information concerning the identity and status of individual trains (like entering or leaving the system). According to an aspect of the present situation, this paper describes an intelligent railroad crossing system. That can be used to avoid collision by road vehicles approaching towards multiple tracks railroad crossing.

161 Design Pattern 1 street traffic (5 in Fig. 1). There is also an audio signalling system (7 in Fig. 1) for unconscious users. The system consists of half barrier gates for escaping trapped vehicles. When there is a train in the system it shuts the gate and opens it after leaving the system. 3. System Block Transmitter Two transmitters (1 & 3 in Fig. 1) are mounted in 3 4 Oscillato r Buffer Amplifi er Power Amplifie r 6 Counter Driver Amplitud e 8 C P c U p 7 5 Fixed Packet Format Fig. 2 Transmitter every train. One is in front and another is back of the train. Each train has a fixed identification number. The train transmits a packet consisting of this identification number along with the position of the transmitter (front or back). The block diagram of the transmitter is shown in Fig. 2. Fig. 1 Top view of the intelligent grade crossing system This automatic railway crossing gate uses radio link for identification, information of approaching and outgoing trains. The train has two transmitters at the beginning (3 in Fig. 1) and end (1in Fig. 1) which transmit an identical packet that can be identified by the sensor (2 in Fig. 1). This packet transmitted through a radio link and received by sensor. Then the sensor sends the information of the packet to cpu (8 in Fig. 1) where the controlling procedure is processed. This cpu consists of several thing like packet identification signalling and gate controller device. After receiving the packet the cpu changes the signal & gate status from the packet type & algorithm stored in the cpu. This system consists of two kinds of signalling posts for safety purpose; one for train (4 in Fig. 1) and another for 3.2 Sensor Sensors (2 in Fig. 1) are situated besides the two railway tracks. Normally, two sensors are placed both side of the railway crossing gate. Sensors receive the packets those are sent by the train s transmitters. Sensors are one kind of radio antenna.

162 Signal 3.5 Central Processing Unit (cpu) Two types of visual signals (4 and 5 in Fig. 1) are used for safety purposes. One is street signal for road traffic and another is train itself. An audio signal is also introduced in the system. Central processing unit (8 in Fig. 1) or central controlling unit connects with three other systems: sensor, gate and signaling. Signal coming from the sensor is processed and decision is sent to the gate and signal. 3.4 Gate The gates (6 in Fig. 1) are consists of mechanical and electrical subsystems. Two half barricade gates are the mechanical portions of the gate. Gates also have electrical portion equipped with driven motor controlled by the cpu. Flag Checker RF Amp Mixer Low Pass Filter Sensor Local Oscillator Serial to Parallel Converter Head Buffer Processing Unit Flag Comparator Tail Checker First portion of this Central processing unit is a radio receiver consists of Radio Frequency Amplifier (RF Amp), Mixer, Low Pass Filter, Sensor, Local Oscillator, Comparator. Then the second part Serial to Parallel Converter with a buffering system built within it. And the processing unit is actually a computer. We used parallel port to input and output data into computer. Visual basic is used for programming the system. The program first take input from parallel port. Then execute the program according working flow of the system. And send out put. 4. Working Flow Diagram The processing unit takes decision according to its own algorithm. The flow chart of decision making algorithm is shown in Fig. 4 where sg_st and sg_tr mention street and train signal status respectively. Both street signal and train signal have two situations {g,r} where g stands for green signal and r for red signal. Here, g_s is the gate status and g_s {o,c} for closing the gate it is c and for opening o. Variable x is a three dimensional array which is actually memory of the system. On the other hand i, j, k are incremental variable and i {1, 2,..., m},where m is total number of memory. The packet has two information; train id (tr_id) and phase (d) of the train. Here, d {h,t} it is h if the packet generate from head of the train otherwise it is t. Variable s indicates at which sensor the signal is acquired from. Variable g_sr {o,c} o when the gate is really open and c when the gate is really close it is actually a feedback status from gate. Street Signal Rail Signal Gate Control Fig 3 Central processing unit

163 142 Start Set: sg_st=g, sg_tr=r, g_s=o, x[] [] []=0 1 Sense Packet? No Yes Pick out tr_id, d, Take no action i=1 x[i][0][0]= =tr_id? Yes k=k+1 k=1 i=i+1 No Yes No i= =m? x[i][0][1]= =s? No x[i][1][0]= =h? Yes Yes No x[k][0][0]= =0? No Yes x[k][0][0]=tr_id x[k][1][0]=d x[k][0][1]=s j=j+1 No j==m? No j=1 x[j][0][0] x[i][0][0] && x[j[0][0] 0? sg_st=r,g_s=c Yes sg_tr=r,g_s=o Yes No g_sr= =c? g_sr= =o? No x[i][0][0]=0 x[i][1][0]=0 x[i][0][1]=0 1 Yes sg_tr=g Yes sg st=g End Fig. 4 Flow chart of the decision making of the system 5. Conclusion An automatic railway crossing gate control system using radio link has been proposed in this paper for reducing the accidents. This system can be implemented both on single and multiple tracks. The main facility of this system is that it can be merged with the existing system. The initial cost of the system is a little high but maintenance cost is very low. Power consumption of the system is low. Using solar energy for providing the power can make the system more fruitful to the rural areas. In future the system would be web based. This system is standalone system and can work 24 7 which is impossible for any man operated system.

164 143 References [1] M. M. Hossain and Sheikh Shanawaz Mostafa, A radiobased railway crossing control system to reduce accident, in Proc. CERIE 2010, 2010, paper C, p. 297 [2] James E. Welk, Railway crossing collision avoidance system, U.S. Patent , Dec. 23, 1997 [3] Keith L. Shirkey, Bruce A. Casella, Wireless train proximity alert system, U. S. Patent , Sep. 10, 1996 [4] Level crossing the free dictionary (2010) The Free Dictionary by Farlex. [Online]. Available: [5] List of level crossing accidents (2010) Wikipedia. [Online]. Available: Sheikh Shanawaz Mostafa completed B.Sc. in Electronics and Communication Engineering Discipline in Khulna University-9208, Khulna, Bangladesh. His current research interest is wireless communication, modulation and biomedical signal processing. His numbers of published papers are four, among them international recognized journal and proceedings of local conference. Md. Mahbub Hossain completed his B.Sc Engineering degree in Electronics and Communication in the year of 2003 from Khulna University, Khulna-9208, Bangladesh. He is now the faculty member of Electronics and Communication Engineering Discipline, Khulna University, Khulna-9208, Bangladesh. His current research interest is wireless communication, modulation techniques, channel coding and fading. His number of published papers are 11 among them international recognized journal and proceedings of international and local conference. Khondker Jahid Reza is serving as a Engineer. He completed his B.Sc. in Electronics and Communication Engineering Discipline in Khulna University, Khulna, Bangladesh. His current research interest is wireless communication, modulation and sensor networks. Gazi Maniur Rashid is currently pursuing B.Sc. in Electronics and Communication Engineering Discipline in Khulna University, Khulna, Bangladesh. His current research interest is wireless communication, modulation.

165 Effective Approaches For Extraction Of Keywords 144 Jasmeen Kaur 1, Vishal Gupta 2 1 ME Research Scholar Computer Science & Engineering, UIET, Panjab University Chandigarh, (UT) Assistant Professor Computer Science & Engineering, UIET, Panjab University Chandigarh,(UT) Abstract Keywords are index terms that contain most important information. Automatic keyword extraction is the task to identify a small set of words, keyphrases or keywords from a document that can describe the meaning of document. Keyword extraction is considered as core technology of all automatic processing for text materials. In this paper, a Survey of Keyword Extraction techniques have been presented that can be applied to extract effective keywords that uniquely identify a document. Keywords:Keyword Extraction, Approaches, Keywords and Document. 1. Introduction Keywords play a crucial role in extracting the correct information as per user requirements. Everyday thousands of books, papers are published which makes it very difficult to go through all the text material,instead there is a need of good information extraction or summarization methods which provide the actual contents of a given document. As such effective keywords are a necessity. Since keyword is the smallest unit which express meaning of entire document, many applications can take advantage of it such as automatic indexing, text summarization, information retrieval, classification, clustering, filtering, cataloging, topic detection and tracking, information visualization, report generation, web searches, etc.[1] Existing methods about Automatic Keyword Extraction can be divided into four categories:- Simple Statistics Approach : These methods are simple and do not need the training data. The statistical information of the words can be used to identify the keywords in the document. Cohen uses N- Gram statistical information to automatically index the document. N-Gram is language and domain independent. Other statistical methods include word frequency, TF*IDF, word co-occurrence, etc[7]. Linguistics Approach : These approaches use the linguistic features of the words mainly sentences and documents. The linguistic approach includes the lexical analysis, syntactic analysis discourse analysis and so on. Machine Learning Approaches : Keyword Extraction can be seen as supervised learning, Machine Learning approach employs the extracted keywords from training documents to learn a model and applies the model to find keywords from new documents. This approach includes Naïve Bayes, Support Vector Machine, etc. Other approaches : Other approaches about keyword extraction mainly combines the methods mentioned above or use some heuristic knowledge in the task of keyword extraction, such as the position, length, layout feature of words, html tags around of the words, etc. Various extraction methods discussed are for single document but these can further applied to multiple documents as per their suitability [12]. Keywords are extracted by identifying the noun phrase as noun phrase comprises very crucial information about the text document. The keywords are choosed based on their linguistic features [2] and informative features [6] such as highlighted words. Query-focused and the words in abstracts or titles can be the part of candidate keywords. Methods such as co-occurrence [8],[14] and machine learning [15],[16] have been used for extracting keywords from a single document. Topics are detected using keyword clustering. In addition extracting and clustering related keywords based on history of query frequency [11] is also one of the methodology adopted. In the following sections various approaches for selecting effective keywords are elaborated. 2. Identify Noun Phrase As Keyword

166 145 Nouns contain bulk of information and this keyword extraction algorithm requires a morphological analyzer and rules or grammar for finding simple noun phrases. Since noun phrases are extracted and become candidate keywords. The noun phrases are scored and clustered and then clusters are scored. The shortest noun phrase from the highest scoring clusters are then used as keywords. The keyword extraction algorithm[1] overview: a) Morphological Analysis Word Segmentation Part of Speech Tagging Stemming b) NounPhrase Extraction and Scoring Noun Phrase Extracted Stopwords Removed Noun Phrase Scored using UnigramFrequency: UF(NP)= NP Σ i=0 UnigramFrequency(Wi) Score(NP)=UF(NP)*NPF(NP) NP Doc Morphological Analysis NP Extraction and Scoring NP Clustering and Scoring Keywords Fig.1. Keyword Extraction Algorithm c) NounPhrase Clustering and Scoring Clusters are formed having a common word Clusters are scored Score(Cluster)= cluster Σ i=0 Score(NP i ) cluster d) Choosing Keywords Shortest length noun phrase choosed as keyword. 3. Term Frequency-Inverse Document Frequency TF-IDF weight evaluates the importance of a word to a document in a collection. Importance increases proportionally to number of times a word appears in document but is offset by frequency of word in corpus [3]. Term T i in particular document D j Term frequency is, tf ij =(n i,j )/ Σ k n k,j where n i,j is number of occurences of considered term(t i ) in document d j and denominator is the sum of number of occurrences of all terms in d j. Inverse Document Frequency[20] is a measure of general importance of term obtained by dividing number of all documents by number of documents containing the term and then taking logarithm of the quotient Idf i =log D {d:t i d} Where, D is total number of documents in corpus and the denominator is number of documents where t i appears. Hence (tf-idf) i,j =tf i,j idf i But the limitation of this method is that it does not work for single document since there are no other documents to compare keywords to algorithms, so it will choose keywords based on term frequency. 4.Selection Based On Informative Features Words are found in various forms of writing in documents which provides additional information about the importance of words[6]. There are various informative features such as, Words emphasized by application of bold, italic or underlined fonts, Words typed or written in upper case, The size of the font applied, Normalized Sentence Length, which is the ratio of number of words occurring in sentence over

167 146 number of words occurring in the longest sentence of the document, Cue-phrases are sentences beginning with summary phrase(in conclusion or in particular) and transition phrase like however, but, yet, nevertheless. 5. Query Focused Keyword Extraction According to this method, keywords correlate to the query sentence and denote the main content of document. It calculates query related feature and then obtains importance of word [9]. The whole system worked as follows, Query Sentence-Pruning Query-Related Feature Keywords are selected The relevant degree of words w1 and w2 is calculated by taking window of length K words. All words in window are said to co-occur with first word with strengths inversely proportional to distance between them. If n(w1,k,w2) is the number of w1 and w2 co-occur in the window, where k denotes real distance between w1 and w2 when they are co-occurred. The relevant degree R(w1,w2) is calculated by, R(w1,w2)= K Σ k=0 w(k)*n(w1,k,w2) Then query-related feature of word w i is, F1(w i )= qwt-1 Σ j=0 R(w i,w j ) 6. Position Weight Algorithm Words in different positions carry different entropy as if same word appears in introduction and conclusion paragraphs, the word carries more information. This Position Weight method record the importance of a word position. It uses three important elements, Paragraph weight Sentence weight Word weight If the paragraph(p j ) is main title or subtitle, leading or concluding paragraph, it carries more weight than common paragraph. First and concluding sentences (s k ) are more important than the example sentences which are weighed nearly zero. Likewise words (w r ) that are capitalized plus some digits are heavily weighed than other common words[2]. The total weight of the term t in document is the sum of weights of all positions it appears. If term t appears m times in document d, its PW is PW(t,d)= m Σ i=1pw (t i ) Where pw of a term in a specific position as Pw(t i ) = pw(t i,p j ).pw(t i,s k ).pw(t i,w r ) For preprocessing, text chunking and elimination of stop wordsthat are included in the Fox stop list[17] have been carried out and leaving special words having transmissible or negative meaning like however, nevertheless and etc. Next is to stem the words using Krovetz algorithm[18] based on WordNet Dictionary[19]. Last is to calculate the PW on the algorithm described. 7. Keyword Extraction Using Conditional Random Field(CRF) Model Conditional random Field (CRF) model works on document specific features. CRF [7] is a state of art sequence labeling method and utilize most of the features of documents sufficiently and effectively for efficient keyword extraction. At the same time, keyword extraction can be considered as string labeling. Here, keyword extraction based on CRF has been discussed. Using CRF model in keyword extraction has not been investigated previously. The results show that CRF model outperforms other machine learning methods such as support vector machines, multiple linear regression model, etc. in the task of keyword extraction. CRF model is a new probabilistic model for segmenting and labeling sequence data. CRF is an undirected graphical model that encodes a conditional probability distribution with a given set of features. In process of manual assignment keyword to a document, the content of document will be analyzed and comprehended firstly. Keywords which can express the meaning of document are then determined. Content analysis is the process that most of the units of a document such as the title, abstract, fulltext, references and so on, be analyzed and comprehended. Sometimes, the entire document has to be read then summarize the content of document, and give the keyword finally. According to process of manual assignment keyword to a document, in this technique, the process is transferred to labeling task of text sequences. In other words, a word or a phrase can be annotated with a label by a large number of features of them. Therefore keyword extraction algorithm based on CRF has been devised to extract keywords. It uses CRF++[13] tool to extract keywords.

168 147 The different kinds of features used are: Local context features such as len, t, a, c, pos, etc. Global context features such as T, A, H, L, R, etc. Process of CRF based keyword extraction:- Preprocessing Docs Trained Feature Feature Extraction CRF Mode Label Fig.2. CRF based Keyword Extraction Process Evaluate Results 7.1 Preprocessing and Features Extraction:- The input is a document. Before CRF model training, transfer the document into the tagging sequences, i.e. a bag of words or phrases of a document. For a new document, the sentence segment has been conducted, and pos tagging Then the features mentioned above are automatically extracted. The output of the feature vectors, and each vector corresponds to a word or a phrase. 7.2 CRF Model Training:- The input is a set of feature vectors by step above. A CRF model has been trained that can label the keyword type. In the CRF model, a word or phrase could be regarded as an example, the keyword has annotated by one kind of labels, such as KW_B, KW_I, KW_S, KW_N, KW_Y. The tagged data are used to training the CRF model in advance. In the CRF++ the output is a CRF model file. 7.3 CRF Labeling and Keyword Extraction:- The input is a document. The document is preprocessed and features are extracted. Then, keyword type has been predicted using CRF model. According to the keyword type, the keywords of the document are extracted. 7.4 Results Evaluation:- Results of keyword extraction can be evaluated by comparing these results with the manual assignment results. 8. Conclusion As with the noun phrase keyword extraction methodology, the only requirement is that the language have a morphological analyzer and rules for finding simple noun phrases. Since nouns contain bulk of the information, noun phrases are extracted and become candidate keywords. The noun phrases are scored and clustered and then the clusters are scored. The shortest noun phrase from the highest scoring clusters are then used as the keywords. The Position Weight algorithm automatically extract keywords from a single document using linguistic features. The results show that the PW algorithm has a great potential for extracting keywords, as it generates a better result than other existing approaches. Using TF-IDF Variants, there are six different values for every word and filtering can be done by using crossdomain comparison i.e. meaningless words have been removed. Furthermore, TTF(Table Term Frequency)[3] has been applied to more precise extraction of keywords. CRF is a state of art sequence labeling method and utilize most of the features of documents sufficiently and effectively for efficient keyword extraction. At the same time, keyword extraction can be considered as string labeling. Here, keyword extraction based on CRF has been discussed. Using CRF model in keyword extraction has not been investigated previously. The results show that CRF model outperforms other machine learning methods such as support vector machines, multiple linear regression model, etc. in the task of keyword extraction. 9. REFERENCES [1] David B. Bracewell and Fuji REN, Multilingual Single Document Keyword Extraction For Information Retrieval, Proceedings of NLP-KE, 2005,pp [2] Xinghua u and Bin Wu, Automatic Keyword Extraction Using Linguistics Features, Sixth IEEE International Conference on Data Mining(ICDMW 06), [3] Sungjick Lee, Han-joon Kim, News Keyword Extraction For Topic Tracking, Fourth International Conference on Networked Computing and Advanced Information Management, 2008,pp [4] Meng Wang, Chao Xu, An approach to Concept-Obtained Text Summarization, Proceedings of IEEE, 2005,pp [5]Tingting He, A Query-Directed Multi - Document Summarization System, Sixth International Conference on Advanced Language Processing And Web Information

169 148 Technology,2007. [6] Rasim M. Alguliev and Ramiz M. Aliguliyev, Effective Summarization Method of Text Documents, Proceedings of International Conference on Web Intelligence, IEEE, [7] Chengzhi Zhang, Automatic Keyword Extraction From Documents Using Conditional Random Fields, Journal of Computational and Information Systems, [8] Y. Matsuo and M. Ishizuka, Keyword Extraction from a Single Document using Word Co-occurrence Statistical Information, International journal on Artificial Intelligence Tools, vol.13, no.1, 2004,pp [9] Liang Ma, Tingting He, Query-focused Multi-document Summarization Using Keyword Extraction, International Conference on Computer Science and Software Engineering, IEEE,2008,pp [10]Christian Wartena, Rogier Brussee, Topic Detection By Clustering Keywords, 19 th International Conference on Database and Expert Systems Application, 2008,pp [11]Toru Onoda, Extracting and Clustering Related keywords based on History of Query Frequency, Second International Symposium on Universal Communication,2008,pp [12] WWW wikipedia.org. [13] CRF++: Yet Another CRFToolkit. [14] Y. Ohsawa, N E Benson, KeyGraph: Automatic indexing by cooccurence graph based on building construction metaphor, In Proceedings of The Advanced Digital Library Conference,1998,vol.12. [15]A. Hulth, Improved Automatic Keyword Extraction Given More Linguistic Knowledge, In Proceeedings of the Conference on Empirical Methods in Natural Language Processing,2003. [16]P Turney, (2000), Learning algorithms for Keyphrase Extraction, Information Retrieval, vol. 2, no. 4, pp [17] C Fox, Lexical Analysis and Stoplists. Information Retrieval: Data Structures and Algorithms, Prentice Hall, New Jersey, 1992, pp [18] R Krovetz., Viewing morphology as an inference process, Proceedings of ACM-SIGIR93,1993,pp [19] G. Miller, Wordnet: An on-line lexical database International Journal of Lexicography,1990,vol. 3,no.4. [20] Stephen Robertson, Understanding Inverse Document Frequency: on theoretical arguments for IDF, Journal of Documentation, Vol. 60, No. 5, 2004, pp AUTHORS INFORMATION Author s Biodata Jasmeen kaur is pursuing ME in Computer Science and Engineering at Unversity Institute of Engineering and Tecnology, Panjab University, Chandigarh. Jasmeen Kaur did her B Tech in CSE from Bhai Gurdas Institute of Engineering and Technology Sangrur in 2007.She secured 80% marks in B Tech. She is carrying out her thesis work in the field of Natural Language Processing. Second Author s Biodata Vishal Gupta is Lecturer in Computer Science & Engineering Department at University Institute of Engineering & Technology, Panjab university Chandigarh. He has done MTech. in computer science & engineering from Punjabi University Patiala in He was among university toppers. He secured 82% Marks in MTech. Vishal did his BTech. in CSE from Govt. Engineering College Ferozepur in He is also pursuing his PhD in Computer Sc & Engg. Vishal is devoting his research work in field of Natural Language processing. He has developed a number of research projects in field of NLP including synonyms detection, automatic question answering and text summarization etc. One of his research paper on Punjabi language text processing was awarded as best research paper by Dr. V. Raja Raman at an International Conference at Panipat. He is also a merit holder in 10 th and 12 th classes of Punjab School education board. in professional societies. The photograph is placed at the top left of the biography. Personal hobbies will be deleted from the biography. [21] Zhang and H Xu, Keyword Extraction Using Support Vector Machines, In Proceedings of Seventh International Conference on Web Age Information Management China, 2006, pp

170 149 Image Splicing Detection Using Inherent Lens Radial Distortion H. R. Chennamma 1, Lalitha Rangarajan 2 1 Department of Studies in Computer Science, University of Mysore Mysore , INDIA 2 Department of Studies in Computer Science, University of Mysore Mysore , INDIA Abstract Image splicing is a common form of image forgery. Such alterations may leave no visual clues of tampering. In recent works camera characteristics consistency across the image has been used to establish the authenticity and integrity of digital images. Such constant camera characteristic properties are inherent from camera manufacturing processes and are unique. The majority of digital cameras are equipped with spherical lens and this introduces radial distortions on images. This aberration is often disturbed and fails to be consistent across the image, when an image is spliced. This paper describes the detection of splicing operation on images by estimating radial distortion from different portions of the image using line-based calibration. For the first time, the detection of image splicing through the verification of consistency of lens radial distortion has been explored in this paper. The conducted experiments demonstrate the efficacy of our proposed approach for the detection of image splicing on both synthetic and real images. Keywords: Image splicing, Lens radial distortion, Straight line fitting, Structural images, Camera calibration. 1. Introduction As digital technology advances, the need for authenticating digital images, validating their content and detection of forgeries has also increased. Common manipulations on images are copying and pasting portions of the image onto the same or another image to create a composite image. Proving the authenticity and integrity of an image is a challenging task. There are two common properties that an untampered image must have: natural scene quality and natural imaging quality [1]. For example, an image with inconsistency between the light direction and the shadow is not authentic because it fails to satisfy natural scene quality. Any output image naturally inherits the characteristic properties of the acquisition device. An image does not meet the natural imaging quality if different parts of the image do not share consistent characteristics of imaging device. A skilled forger can manage to satisfy the natural scene quality by using sophisticated image editing software but natural imaging quality is very difficult to achieve. This motivates us to take up this research. The aim of this research is to demonstrate that it is possible to use inherent lens aberrations as unique fingerprints in the images for the detection of image splicing. Inconsistency in the degree of lens distortion in different portions of the image leads to the detection of image tampering. It is generally accepted that the optics of most consumer level cameras deviate from the ideal pinhole camera model. Among different kinds of aberrations, lens radial distortion is the most severe. The lens radial distortion causes non-linear geometrical distortion on images. In this paper we propose a novel passive technique (with no watermark or signature) for detecting copy-paste forgery by quantitatively measuring lens radial distortion across the image using line-based calibration. We estimate lens radial distortion from straight edges projected on images. Hence neither calibration pattern nor information about other camera parameter is necessary. Usually a composite image is created not just by weaving different portions of the same or different images, but it is also accompanied by subsequent operations like JPEG compression, contrast/brightness adjustment, color changing, blurring, rotation, resizing etc. to hide the obvious traces of tampering. The remainder of this paper is organized as follows. Section 2 reviews the relevant research on image splicing detection. Section 3 details the background of lens radial distortion and also describes how to estimate lens radial distortion parameter for each detected line segment in the image. Section 4 analyses the effect of lens distortion at

171 150 various zoom levels, for different cameras and also presents experimental results on the detection of image tampering performed on both synthetic images and real images. Section 5 discusses the limitations of the proposed method and future work. Section 6 concludes the paper. 2. Related Work Digital image forensics is emerging as an interesting and challenging field of research [2, 3] recently. Here we present the review of related works on image splicing detection. One of the kinds of image tampering is object removal where the regions of unwanted objects in an image are replaced by other parts of the same image. This type of operation is called copy-move or region-duplication. Methods in [4-6] are specifically designed to detect region duplication and are all based on block matching. First, the method divides an image into small blocks. Then it extracts the features from each block and hence, identifies possible duplicated regions on comparison. The main difference of these methods is the choice of features. Fridrich, et al. [4] have analyzed the DCT coefficients from each block. Popescu, et al. [5] have employed the principal component analysis (PCA) to reduce the image blocks into a PCA feature vector. Luo et al. [6] have extracted seven features in each block. Their experimental results demonstrated that the method could resist more post-processing operations. Another kind of image tampering is splicing. Unlike region duplication, image splicing is defined as a simple joining of fragments of two or more different images. Several researchers have investigated the problem of splicing based on statistical properties of pixels (called pixel-based techniques) and camera characteristics (called camera-based techniques). Now, let us briefly review the literature on both techniques. Johnson et al. [7] have described a method for revealing, traces of tampering using light inconsistencies as it is often difficult to match the lighting conditions from the individual photographs. Tian-Tsong Ng et al. [8] have described techniques to detect photomontaging. They have designed a classifier based on bi-coherence features of the natural images and photomontaged images. They have also proposed a mathematical model for image splicing [9]. One of the fundamental operations that need to be carried out in order to create forgeries is resizing (resampling). It is an operation that is likely to be carried out irrespective of the kind of forgery (copy move, photomontage, etc). Farid et al. [11] have described a method for estimation of resampling parameters in a discrete sequence and have shown its applications to image forensics. Chen et al. [12] have analyzed phase congruency for detection of image splicing. We have explored the use of wavelets for the detection of resampled portions [13]. Methods for the detection of image alteration based on inherent characteristics of digital camera (camera-based techniques) have been reported in [14-18]. Johnson et al. [14] have explored lateral chromatic aberration as a tool for detecting image tampering. Lateral chromatic aberration manifests itself, to a first-order approximation, as an expansion/contraction of color channels with respect to one another. When tampering with an image, this aberration is often disturbed and fails to be consistent across the image. As the authors mentioned, this approach is effective only when the manipulated region is relatively small allowing for a reliable global estimate. Copy-paste forgery in JPEG images has been detected by extracting the DCT block artifact grid and by identifying mismatch among the grid by Weihai Li et al. [10]. Farid et al. [15] have noticed that the color images taken from a digital camera have specific kind of correlations among the pixels, due to interpolation in the color filter array (CFA). These correlations are likely to be destroyed, when an image is tampered. They have showed that the method can reasonably distinguish between CFA and non-cfa interpolated portions of images even when the images are subjected to JPEG compression, additive noise or luminance non-linearities. But they have not discussed splice detection, when portions of different images with same CFA interpolation technique are spliced together as a composite image. Lukas et al. [16] have presented an automatic approach for the detection of tampered regions based on pattern noise, a unique stochastic characteristic of imaging sensors. The regions that lack the pattern noise are highly suspected to be forgeries. The method works in the presence of either the camera that took the image or when sufficiently many images taken by that camera are available. However this is always not possible. A semiautomatic method for the detection of image splicing based on geometry invariants and camera characteristic consistency have been proposed by Hsu [17]. The method detects Camera Response Function (CRF) for each region in an image based on geometry invariants and subsequently checks whether the CRFs are consistent with each other using cross-fitting techniques. CRF inconsistency implies splicing. The authors have used only uncompressed RAW or BMP image formats which are not always provided with all consumer level compact digital cameras, whereas our proposed approach is not restricted to the type of image format. Johnson et al. [18] have described a technique specifically designed to detect composites of images of people. This approach estimates a

172 151 camera s principal point from the image of a person s eyes. Inconsistencies in the principal point are then used as evidence of tampering. As authors mentioned, the major sensitivity with this technique is in extracting the elliptical boundary of the eye. This process will be particularly difficult for low-resolution images. Though the method proposed in this paper uses camera characteristic property; lens radial distortion for splicing detection, the method can successfully detect both copymove and copy-paste forgery, even if they are created by using images of the same camera. 3. Lens Radial Distortion Virtually all optical imaging systems introduce a variety of aberrations into an image due to its imperfections and artifacts. Lens distortion is one such aberration introduced due to geometry of camera lenses. 3.1 Background of lens radial distortion Unlike extrinsic factors, intrinsic factors are due to camera characteristics and are specific constants to a camera, e.g. focal length, imaging plane position and orientation, lens distortion, aspect ratio etc. These are independent of position and nature of the objects captured. Lens distortion is deviation from rectilinear projection; a projection in which straight lines in a scene remain straight in an image. However, in reality almost all lenses suffer from small or large amounts of distortion. Lens radial distortion is the dominating source of mapping errors especially in inexpensive wide-angle lenses because wide-angle lens is shaped to allow a larger field of view. Due to the shape of the lens magnification and focus is not isotropic resulting in unwanted distortions. Lens radial distortion deforms the whole image by rendering straight lines in the object space as curved lines on the film or camera sensor. Radial distortion is a non-linear transformation of the image increasing from the centre of distortion to the periphery of the image. The centre of lens distortion is a point somewhere close to the centre of the image, around which distortion due to the shape of the camera lens is approximately symmetrical. Fig. 1 shows an image of a grid and r is the radius of grid. Two of the most common distortions are barrel (k>0) and pincushion (k<0) distortions (shown in fig. 2a & 2b), where k is the distortion parameter which indicates the amount of lens radial distortion (refer Section 3.2). r` in fig. 2 is the deformed radii of the grid due to distortions. Fig. 1 No distortion (a) Barrel distortion (b) Pincushion distortion Fig. 2 Distorted images It is evident from fig. 2a & 2b that the amount of distortion increases with distance from centre of image to the periphery of image. We have found that this correlation among different portions of the image is disturbed in spliced images. In order to prove the integrity of an image, in this paper, we look for the consistency among the lens radial distortion parameters which are estimated from different regions of an image. This source of information is very strong, provided 3D lines are projected on the image. Thus the proposed technique works on images of city scenes, interior scenes, aerial views containing buildings and man-made structures. Apart from the lens design, the degree of radial distortion is related to focal length [19]. Usually, lenses with short focal length have a larger degree of barrel distortion, whereas lenses with long focal length suffer more from pincushion distortion. As a result, lenses from different camera leave unique imprints on the pictures being captured. 3.2 Measuring radial distortion of lenses The lens radial distortion model can be written as an infinite series, as given below: 2 4 r r (1 k r k r ) (1) u with r u d 1 d 2 d 2 2 u yu rd x and x 2 d y r u and rd are undistorted and distorted radii respectively. 2 2 Radius is the radial distance x y of a point x, y from the centre of distortion. From (1) it follows that: 2 4 x u xd (1 k1rd k2rd ) (2) 2 4 y u yd (1 k1rd k2rd ) (3) Centre of distortion is taken as centre of image [19]. The first-order radial symmetric distortion parameter k 1 is sufficient for reasonable accuracy [20]. Thus the polynomial distortion model in (2) and (3) may be simplified as 2 x x (1 k r ) (4) u u d d 1 d 2 y y (1 k r ) (5) 1 d 2 d

173 152 The parameter k 1 has dominant influence on the kind of radial lens distortion. If k 1 >0, distortion is barrel and if k <0, distortion is pincushion Proposed approach Portions of an image are correlated with each other with respect to the imaging device. Such correlations will be disturbed in spliced images. An intrinsic camera parameter viz., lens radial distortion is used for the detection of image splicing. Inconsistency in the degree of lens radial distortion across the image is the main evidence for splicing operation. In this section, a novel passive technique is described for detecting copy-paste forgery by quantitatively measuring lens radial distortion from different portions of the image using line-based calibration. Line-based calibration of lens radial distortion can be divided into three steps: Detection of edges with sub-pixel accuracy Extraction of distorted line segments Estimation of k 1 for each distorted line in an image Detection of edges with sub-pixel accuracy: The first step of the calibration consists of detecting edges from an image. Since image distortion is sometimes less than a pixel, there is a need for an edge detection method with sub-pixel accuracy. We used the edge detection method proposed in [21], which is a sub-pixel refinement of the classic non-maxima suppression of the gradient norm in the direction of the gradient. Extraction of distorted line segments: This calibration method relies on the constraint that straight lines in 3D must always project as straight lines in the 2D image plane, if the radial lens distortion is compensated. In order to calibrate distortion, we must find edges in the image which are most probably projections of 3D segments. Because of the distortion, a long segment may be broken into smaller segments. By defining a very small tolerance region, we can extract such distorted line segments as shown in fig. 3. Perturbations along the lines may be generated due to low resolution of the image. Such distracted or perturbed line segments must be rejected even within the tolerance region. Estimation of k 1 for each distorted line in an image: Measure the degree of distortion in terms of distortion parameter k1 for each distorted line in the image. In order to measure the absolute deviation of a distorted line from its undistorted line, the points on a distorted line segment are used to fit a straight line using linear regression [22]. From eq. (4) and eq. (5), all n distorted points p d,i =(x d,i y d,i ) of the selected curved line are mapped to the undistorted points p u,i =(x u,i y u,i ) where (1 i n) as follows: 2 xu, i xd, i (1 k1rd, i ) (6) 2 yu, i yd, i (1 k1rd, i ) (7) All n undistorted points p u,i should now lie on a straight line. Thus, an associated straight line L n is represented in Hesse s normal form, it has three unknowns n x,n y and d 0 : T nx x L : 0 n d (8) n y y To determine these unknowns with linear regression, the following expressions are calculated. n n 1 1 X x d, i Y y d, i n n i1 n x d, i i1 2 2 X Y 1 n i1 n i1 n XY 1 x d, i yd, i n i1 Two cases have to be distinguished: 2 1 n 2 y d, i Case 1: Let 2 X X Y Y, then the associated straight line L n is parameterized as L n : y ax b (9) with a XY X 2 X Y X 2 b X 2 X Y X 2 X 2 XY Figure. 3 Detection of broken line segments within the tolerance region and the three unknowns are as follows: a 1 n x n y d a 2 1 a a b 2 1

174 Case 2: Let 2 X X Y Y, then the parameterization of the associated straight line L n changes to: L n : x cy d (10) with c XY Y 2 X Y Y 2 d X 2 Y Y X In this case the three unknowns of eq. (8) are 1 c n x n y d c 2 1 c Y 2 c XY d 2 1 Now an associated straight line L n is found, which is a function of k1 and the points p d,i. Thus a cost function with the residual errors n of eq. (8) is formulated as: n i n x y T x y u, i u, i d Substitute x u,n and y u,n from eq. (6) and (7) n i n x y T x y d, i d, i (1 k r 2 1 d, i (1 k r 2 1 d, i n 0 ) d ) Where k1 is selected to minimize 2 i. i1 This cost function is a non-linear function of x d,i and y d,i of a curved line or distorted line. The deviation of points (x d,i y d,i ) from their original positions (x u,i y u,i ) is used to estimate the amount of distortion. The distortion error is the sum of squares of distances from the points to the straight line. To estimate distortion parameter k 1 the sum of squares is minimized using the iterative Levenberg- Marquardt method through lsqnonlin function found in MATLAB. Thus we obtain unique k 1 for each distorted line segment in an image depending on the amount of distortion. 4. Experimental Setup Three sets of experiments were performed. The first set of experiments aims at technical investigation and calibration of lens radial distortion for different consumer level compact digital cameras. The second set of experiments conducted on synthetic images to show how to use radial 0 distortion parameter as a feature to detect image splicing. The third set of experiments study the performance of proposed features on real images. 4.1 Analysis of lens radial distortion for different cameras To analyze the behavior of intrinsic radial distortion parameter across the image, we have used 6 digital cameras of recent models from four different manufacturers. The configurations of the cameras are given in table 1. Table 1. Cameras used in experiments and their properties Camera Brand Focal Length(mm) Optical Zoom Resolution (Mega Pixel) Sony DSC-W x 7.2 Sony DSC-W x 12.1 Canon A x 7.1 Casio EX-Z x 7.1 Nikon S x 10 Sony DSC-W x 10.1 Note: The focal length is equivalent to a 35mm film camera Fig. 4 shows the checker-board with 9 by 12 square grids, generated manually without any distortions. Fig. 5 shows the extracted straight lines from fig. 4. The lens radial distortion parameter k 1 is computed for each straight line (ref section 3.3) and is reported zero for all straight lines and the same is shown as a graph in figure 6. All through the experiments, it is assumed that the image centre as (0,0) and the horizontal and vertical coordinates are normalized so that the maximum of the dimensions is in the range (-1,1). Most consumer digital cameras are equipped with an optical zoom lens. The lens radial distortion parameters change with focal length, which usually varies from barrel at the wide end to pincushion at the tele end. In this section we investigate the impact of optical zoom on the behaviour of radial distortion parameter across the image. To study the actual behaviour of radial distortion parameter k 1 across the image, images of the same scene are acquired from different cameras. The checker board (shown in fig. 4) is captured, approximately with same distance, position and orientation of the camera. The images were taken with no flash, auto-focus, JPEG format and other default settings. The sample images captured by Sony DSC-W35 camera with different zoom levels is shown in fig. 7. Fig. 8 is the corresponding edge images of each image in fig. 7. Since the radial distortion is approximately symmetric we have drawn the graph of k 1 for vertical lines which lies in the right of image centre. Figures 9-14 show the behaviour of radial distortion parameter k 1 across the image for various cameras at

175 154 different zoom levels. It is clear that no camera is ideal. All cameras have noticeable amount of radial distortion. It is also evident from the graph that the degree of radial distortion increases with the distance from centre of the image. We can also observe that the radial distortion parameter changes with zoom and most of the cameras vary from barrel at the wide end to pincushion at the tele end (refer fig. 9, 10, 12 and 14). Fig. 13 Fig. 14 Figs Behaviour of lens radial distortion parameter k 1 across the image for various cameras at different zoom levels 4.2 Experiments on Synthetic Images Fig. 4 Checker-board Fig. 5 Extracted edges Fig. 6 Graph of k 1 for lines in fig. 5 Fig. 7 Checker-board image captured by Sony DSC-W35 camera Fig. 8 Extracted edges from fig 7 Fig. 9 Fig. 10 Fig. 11 Fig. 12 In this section we describe, how to use lens radial distortion parameter for the detection of image splicing. If two image regions belong to the same image (untampered) then the radial distortion of extracted line segments should behave as expected (explained in Section 4.1). That is lines that are more or less equidistant will suffer from more or less equal amount of radial distortion. Also radial distortion is uniform, barrel or pincushion. Usually the composites are created in three ways: (i) Image splicing on the same image (copy-move forgery) (ii) Two or more images of the same camera (may be captured at different zoom) are used to create a composite image (iii) Two or more images of the different cameras (irrespective of the camera make or model) are used to create a spliced image Hence, we used three different cameras, in which two are from same manufacturer, so that all types of splicing described above can be carried out. Fig. 15 is the image acquired from Sony DSC-W35 and the extracted straight lines are shown in fig. 16. Table 2 shows the distances (column 2) of straight lines from the image centre and its corresponding values of k 1 (column 3). Line 1 is the left most and line 6 is the right most line. You can observe that the values of k 1 gradually increase from image centre to its periphery. Similar observations on k 1 were already noticed in Section 4.1 (refer graph of k 1 values of various cameras in figures 9-14). This consistency will be disturbed in case of copy-move or copy-paste forgery. Inconsistency in lens radial distortion can be detected if one of the two conditions is not met: (a) Amount of radial distortion is symmetric and increases with the distance from centre of the image and (b) Sign of k 1 should be same for all lines throughout the image. In the untampered image (fig. 15), all bottles are of different colours. The composites were created by copying and pasting a bottle over another bottle. The different cases of image splicing described in (i) to (iii) are shown in figure The original images were captured at different zoom levels and some post-processing operations

176 155 like rotation and scaling have also been done while pasting a portion. The values of k 1 for each line from left to right for images in figure are listed in tables 3-6 respectively. Inconsistency is highlighted in all the tables. When a bottle on right is pasted over a bottle on left (ssown in fig. 17) then the radial distortion of the corresponding edges would change to barrel (pincushion) where as the actual distortion in untampered image would have been pincushion (barrel). This can be noticed in table 3. Image in figure 18 has been spliced by replacing the middle bottle by a bottle on the right side of another image and k 1 values for all extracted lines is given in table 4. k 1 is negative for line 3 and positive for all other lines indicating inconsistency of type (b). Observe that k 1 of line 4 is greater than that of line 5, this implies inconsistency of type (a). Thus splicing has been successfully detected. robust to resizing and blurring provided the extracted lines continue to be unperturbed. Fig. 17 Copy-move or Regionduplication Fig. 18 Composite of 2 images captured by cameras of different make Fig. 19 Composite of two images Fig. 20 Composite of two images captured by same camera captured by cameras of different model Fig. 15 An image of Sony DSC-W35 Fig. 16 Extracted Edges Table 2. Estimated k 1 for each line (from left to right) of fig. 16 Straight line no. Distance from centre Distortion parameter k In real cases, the creation of composites is commonly accompanied with subsequent operations like JPEG compression, contrast/brightness adjustment, color changing, blurring, rotation, resizing etc. to hide the obvious traces of tampering. Since the proposed splicing detection method works on images consisting of straight edges, contrast, brightness and color manipulations will not affect edge detection, unless the object color and the background color are indistinguishable. It is evident from the above experiments that the proposed approach is robust to rotation, as distortion is same even if lines are rotated. We observed that JPEG compression, blurring and resizing operations affects on the performance of the proposed method. Experiments show that JPEG compression with a quality factor of 5 (out of 10) also can detect straight lines with reasonable accuracy sufficient to estimate k 1. Resizing is the most common operation performed while creating composites, in order to match with the size of the host image. The proposed approach is Table 3. Estimated k 1 for each line (from left to right) in fig. 17 Straight line no. Distance from centre Distortion parameter k Table 4. Estimated k 1 for each line (from left to right) in fig. 18 Straight line no. Distance from centre Distortion parameter k Table 5. Estimated k 1 for each line (from left to right) in fig. 19 Straight line no. Distance from centre Distortion parameter k Table 6. Estimated k 1 for each line (from left to right) in fig. 20 Straight line no. Distance from centre Distortion parameter k

177 Experiments on real images The non-availability of the suitable data set for examining the proposed method led us to create our own Spliced Image Data Set (SIDS). However, we have also experimented with a few images from the database provided by Hsu [17]. In order to compute the splice detection rate, we have created 100 spliced images from 350 authentic images. Authentic images were taken with our 6 consumer level compact digital cameras (mentioned in Table 1) and 50 images are downloaded from internet. 50 images were captured from each of 6 cameras in JPEG format with dimensions ranging from 1632x1224 to 3072x2304. These images mainly contain indoor scenes like computers, boards, tables, library, photo frames etc. Some images contain outdoor scenes like buildings, shopping complex etc. We created spliced images from the authentic image set using Adobe Photoshop. In order to hide the traces of tampering, some post processing operations like resizing and rotation were performed. (c) (d) (i) (j) As a first step, all distorted line segments from each spliced image are detected as described in Section 3.3. Further, for accurate comparison of distortion parameter k for each detected line in an image, the length of those 1 lines must be equal. All lines detected in an image are trimmed to that of the shortest line (at least 1/3 rd of the image height). The detection rate for our spliced image dataset is found as 86%. Some sample spliced images of our dataset and their extracted line segments are shown in fig. 21. The pasted portion is indicated by red lines. (a) (g) (e) (f) Fig. 21. (a)-(f) Sample spliced images and (g)-(l) are their extracted edges respectively We have also verified our proposed approach with images selected from Columbia Image Splicing Detection Evaluation dataset. However, we have selected those spliced images that contain straight edges. Some sample images are shown in fig. 22. The pasted portion is indicated by red lines. (k) (l) (a) (d) (b) (h)

178 157 (b) (e) minimum length of 1/3 rd of the image height from different regions in order to prove the integrity of the image. To summarize, the proposed method works best on images with lot of lines and of significant lengths, such as images of city scenes, interior scenes, aerial views containing buildings and man-made structures. Probably the experimental results can be improved by using a more sophisticated method to estimate the lens radial distortion across the image. We expect the proposed technique, when integrated with other available splicing detection methods [14-18], to become more effective in exploring digital forgeries. (c) (f) Figs. 22. (a)-(c) Sample images from database [17] and (d)-(f) are their extracted edges respectively 5. Discussion and future work The proposed approach detects splicing of images when straight edges are available, but the main difficulty lies in the extraction of such straight edges. Sometimes the low quality image will generate perturbations along the straight lines which results in wrong estimation of radial distortion parameter k 1. Because of low resolution, an authentic image in fig. 23 is detected as spliced image (perturbations may be visible in enlarged version). (a) (b) Fig. 23. (a) A low quality image and (b) shows the extracted perturbed lines We point out that our method may not provide sufficiently conclusive statistical evidence regarding spliced portions of the image. It means the proposed approach may fail to detect a spliced portion if two images have same kind of distortion are spliced together and the position of copied and pasted portion is same with respect to the centre of image. Also, more research and analysis are needed to determine the influence of centre of distortion and the length of lines. Though the proposed method works for the images from lower-end consumer level digital cameras taken with default settings and most commonly available JPEG image format, we need at least two or more straight lines with the It is interesting to address malicious attacks intended to interfere the detection algorithm, such as adding or removing radial distortion from an image [23, 24]. Since it is possible to manipulate the distortion parameters of an image globally, these alterations will not affect the performance of our method if done on spliced image. 6. Conclusion Portions of the image are correlated with each other with respect to the imaging device. Such correlations will be disturbed in spliced images. We have used an intrinsic camera parameter, namely lens radial distortion, for the detection of image splicing. Inconsistency in the degree of lens radial distortion across the image is the main evidence for the detection of spliced images. In this paper we propose a novel passive technique (with no watermark or signature) for detecting copy-paste forgery by quantitatively measuring lens radial distortion from different portions of the image using line-based calibration. Experiment in section 4.1 shows that most consumer level digital cameras have small or large amount of lens radial distortion at different zoom levels. Experimental set up in section 4.2 demonstrates how efficiently the lens radial distortion parameter k 1 may be used for the detection of image splicing and the experimental results in section 4.3 shows that our method works well in case of real images. The primary contribution of our work is that we examine the use of inherent lens distortion as a unique imprint on the images for the detection of image splicing. References [1]. T.-T. Ng, S.-F. Chang. J. Hsu, L. Xie, M.-P. Tsui, Physics motivated features for distinguishing photographic images and computer graphics, Proc. 13 th ACM Internat. Conf. on Multimedia, Singapore, 6-11 November 2005, pp [2]. H. Farid, A survey of image forgery detection, IEEE Signal

179 158 Processing Magazine, 26(2), (2009), [3] Luo Weiqi, Qu Zhenhua, Pan Feng, Huang Jiwu, A survey of passive technology for digital image forensics, Front. Computer Science China, 1(2), (2007), [4]. J. Fridrich, D. Soukal, J. Lukas, Detection of copy-move forgery in digital images, Proc. DFRWS, USA, [5]. A.C. Popescu and H. Farid, Exposing digital forgeries by detecting duplicated image regions, Tech. Rep. TR , Dept. of Computer Science, Dartmouth College, [6]. W. Luo, J. Huang, G. Qiu, Robust detection of regionduplication forgery in digital image, Proc. 18 th Intl. Conf. on Pattern Recognition, Vol. 4, 2006, pp [7]. M.K. Johnson and H. Farid, Exposing digital forgeries by detecting inconsistencies in lighting, Proc. 7 th Workshop on ACM Multimedia and Security, New York, 2005, pp [8]. T.-T. Ng, S.-F. Chang, Q. Sun, Blind detection of photomontage using higher order statistics, Proc. IEEE Intl. Symposium on Circuits and Systems, 2004, [9]. T.-T. Ng, and S.-F. Chang, A model for image splicing, Proc. IEEE Internat. Conf. on Image Processing, Singapore, Vol. 2, Oct. 2004, pp [10]. Weihai Li, Yuan Yuan, Nenghai Yu, Passive detection of doctored JPEG image via block artifact grid extraction, Signal Processing 89(9), (2009), [11]. H. Farid and A.C. Popescu, Exposing digital forgeries by detecting traces of re-sampling, IEEE Transactions on Signal Processing 53(2), (2005), [12]. W. Chen, Y.Q. Shi, W. Su, Image splicing detection using 2-D phase congruency and statistical moments of characteristic function, Proc. of SPIE: Security, Steganography, and Watermarking of Multimedia Contents IX, CA,Vol. 6072, 2007, pp 65050R R-8. [13]. H.R. Chennamma, Lalitha Rangarajan, Ch. Gandhi, Narinder Singh, Detecting Forgeries for Digital image Forensics. Journal of Indian Academy of Forensic Sciences, Vol 41, (2007), [14]. M. K. Johnson, Hany Farid, Exposing digital forgeries through chromatic aberration, Proceedings of the 8 th workshop on Multimedia and security, 2006, Geneva. [15]. H. Farid and A.C. Popescu, Exposing digital forgeries in color filter array interpolated images, IEEE Transactions on Signal Processing 53(10), (2005), [16]. J. Lukas, J. Fridrich, M. Goljan, Detecting digital image forgeries using sensor pattern noise, Proc. of SPIE Electronic Imaging, Security, Steganography, and Watermarking of Multimedia Contents VIII, San Jose, CA, Vol. 6072, Jan. 2006, pp. 0Y1-0Y11. [17]. Y.F. Hsu and S.F. Chang, Detecting image splicing using geometry invariants and camera characteristics consistency, Proc. IEEE Internat. Conf. on Multimedia and Expo., Toronto, Canada, 9-12 July 2006, pp [18]. M.K. Johnson and H. Farid, Detecting photographic composites of people, Proc. 6 th International Workshop on Digital Watermarking, China, Vol. 5041, 2008, pp [19]. K.S. Choi, E.Y. Lam, K.Y. Wong, Automatic source camera identification using the intrinsic lens radial distortion, Optics Express 14(24), (2006), [20]. F. Deverney and O. Faugeras, straight lines have to be straight, Machine vision and appln. 13(1), (2001), [21]. F. Deverney, A non-maxima suppression method for edge detection with sub-pixel accuracy, Technical Report RR- 2724, INRIA, (1995). [22]. T. Thormählen, H. Broszio, I. Wassermann, Robust Line- Based Calibration of Lens Distortion from a Single View, Proc. of Mirage, INRIA Rocquencourt, France, 2003, pp [23]. H. Farid and Alin C. Popescu, Blind removal of lens distortion, Journal of the Optical Society of America A, 18(9), (2001), [24]. J. Weng, Camera calibration with distortion models and accuracy evaluation, IEEE Trans. on Pattern Analysis and Machine Intelligence, 14(10), (1992), H.R. Chennamma received her graduate degree in Computer Applications with distinction in the year 2003, Vishwesharaiah Technological University, India. She was a Project Trainee for a year at the National Aerospace Laboratory (NAL), Bangalore, India. She served as a software engineer for a year in a multinational software company, Bangalore, India. Mrs. Chennamma is now a Senior Research Fellow (SRF) in National Computer Forensic Laboratory, Ministry of Home Affairs, Government of India, Hyderabad since Subsequently, she was awarded Ph.D. program fellowship at the Department of Computer Science, University of Mysore in the year Mrs. Chennamma is the recipient of Best Scientific Paper Award in the All India Forensic Science Conference, Kolkata, India in the year Her current research interests are Image Forensics, Pattern Recognition and Image Retrieval. Dr. Lalitha Rangarajan has two graduate degrees to her credit, one in Mathematics from University of Madras, India (1980) and the other in Industrial Engineering (Specialization: Operational Research) from Pardue University, USA during (1988). She has taught Mathematics briefly at graduate level, for 5 years during 1980 to 1985 in India. She joined Department of Computer Science, University of Mysore, to teach graduate students of the Department, in 1988, where she is currently working as a Reader. She completed Ph.D. in Computer Science in the area of pattern recognition in She is presently working in the areas of Feature Reduction, Image Retrieval and Bioinformatics.

180 159 Designing A Re-Configurable Fractional Fourier Transform Architecture Using Systolic Array Anal Acharya 1 and Soumen Mukherjee 2 1 Dept. of Computer Science, St. Xavier s College, Kolkata, West Bengal , India. 2 Dept. of Computer Application, RCC Institute of Information Technoogy, Kolkata, West Bengal , India. Abstract FRFT (Fractional Fourier Transforms) algorithm, which has been derived from DFT, computes the angular domains within the time and frequency domains. This algorithm is increasingly used in the field of signal filtering, quantum mechanics and optical physics. In this paper we develop an efficient, systolic, reconfigurable architecture for a particular type of FRFT called MA-CDFRFT (Multi Angle Centered Discrete FRFT). The benefit of this particular type of FRFT is that it computes all the signal components within equally spaced angles. Systolic architecture is used for this computation as it has certain advantages over the other forms like simplicity, regularity, concurrency and computation intensive The resultant product so developed should meet the challenges of today s market like marketable and cheap along with meeting customer demands. This calls for the architecture to be re-configurable. Reconfigurable computer consist of a standard processor and an array of re-configurable hardware. The main processor would control the behavior of the re-configurable hardware. The reconfigurable hardware would then be tailored to perform a specific task, such as image processing or pattern matching applications, as if it was built to perform this task exclusively. Keywords: MA-CDFRFT, Systolic Array, Up/down array, Reconfigurable PE. 1. Introduction The DFT algorithm has been replaced by FFT algorithm by the signal processing researchers for its lower computational complexity. Also DCT and DWT algorithms are finding increasing importance in the field of signal compression. DFT had only one basic definition and a variety of algorithms have been devised for its fast computation. But when FRFT is analyzed in discrete domain, there are many definitions of discrete fractional Fourier transform (DFRFT) [3]. We first define Centered DFRFT (CDFRFT) and extend this definition to Multi-Angle CDFRFT (MA-CDFRFT). All through we shall use the definition given by [1]. Our proposed architecture can handle real time data and has reduced computational complexity using systolic up/down array in FFT computation. We then propose a re-configurable architecture for FRFT. The last of these steps corresponds to FFT implementation. Finally we construct a re-configurable PE, which will work for each stage. The various stages of the PE is generated by a set of signals from the control unit. 2. Related works An architecture of FRFT has been proposed by Sinha et. al. [16] but that method is not suitable for real time data. Dick has proposed a method of computing DFT on FPGA based Systolic Arrays [8]. Dick also proposed a method for computing multi-dimensional DFT using Xilinx FPGA [17]. Cho et.al. discussed a implementation of DCT algorithm for parallel architecture[18]. A re-configurable architecture has been discussed by Acharya et. al.[19][20]. 3. Computation of MA-CDFRFT CDFRFT can be expressed [1] using Equation 1 as follows N-1 {A α } kn = V kp V np e -jpα (1) p=0 Where V kp is the k-th element of eigenvector p. Multiplying A α by the signal element x[n] and rearranging, we obtain

181 160 N-1 N-1 X α [k] = V kp x[n] V np e -jpα (3) p=0 n=0 For a set of equally spaced values of a given by: α r = 2πr/N r=0,1,,n-1 (4) that correspond to the cases for which the trace of Aa becomes zero, we can rewrite the transform in terms of index r as N-1 N-1 X α [k] = V kp x[n] V np e -j(2π/n)pr (5) p=0 n=0 For the ease of computation we define Z + k [p] as N-1 Z + k [p]= x[n] V np (6) n=0 Fig 1. Loading of Constants V np In the second step (Figure 3) each of the signal elements are multiplied with the Eigen vector element Vnp. In the first cycle V 00, V 10,, V n0 is multiplied by x[0], x[1],.., x[n-1] respectively. In the next cycle these multiplied values move to the second row whereas the first row multiplies the next set of signal elements with V 11,, V n1. Finally in cycle N the value derived from each of the processing element are added at and the value Z + k [0] is derived. Again defining Z k [p] as N-1 Z k [p]= V kp x[n] V np (7) n=0 we can see that the transform can be expressed as a DFT, that is N-1 X α [k] = Z k [p] e -j(2π/n)pr (8) p=0 where r=0,1,,n-1 and k=0,1,,n-1. Since Xk[r] contains all the CDFRFTs corresponding to the discrete set of angles ar. [1] suggested that this matrix be called Multi-angle-CDFRFT or MA-CDFRFT. 4. Proposed FRFT Architecture In the first step (Figure 1) the elements of the Eigen vector element Vnp are loaded into the processing elements [20]. Fig 2. Multiplying each single element by V np & taking the sum We now discuss the third step (Figure 5) of FRFT architecture [20].. The elements derived at (the adder) of step 2 of FRFT are transferred systolically to the processing elements containing the elements containing V 00, V 01,, V N-10. Thus the elements Z k [0], Z k [1],..,Z k [N-1] are derived(figure 5).

182 161 Fig 3. Calculation of Z k [P] Fig 4: Detailed Working of the up/down array in FFT architecture This element multiplied by the twiddle factor summed from 0 to N-1 gives the corresponding FRFT component for the rth plane. This computation can be done in a fast manner using FFT. We propose this computation be done using a UP/DOWN systolic array [19], which essentially consists of one upwards path for the bottom N/2, sets of input data and a downward path for the top N/2 sets of input data. 5. Computational Complexity of the Architecture First we define Utilization Factor: Let there be N number of processing elements. In a particular cycle i, if N i be the number of processing elements used (N i <=N) then the utilization factor at cycle i is defined as U i = N i /N We divide the computation into the following stages: In the first stage v np is loaded in the systolic array. We assume there are N cells which stores V 0p to V N-1p and T 1 is the transfer time of the signal element from the left to the right cell. Then the time needed to fill the two dimensional systolic array is NT 1. Here the utilization factor of the systolic array after NT 1 time is 100 percent. In the second stage computation of x[n] v np is done. We assume that clock period of real addition is T 2 and clock period to switch between real to imaginary signal components the selector switch [20] is T 3 then the total time is: N (N-1)/2*T 1 + T 2 log 2 N + T 3. The PE utilization is 50% in this stage. In the third stage f v kp x[n] v np is computed. Again we assume the time for real multiplication and real addition is T 4 and to configure the required connection is T 5. So the total time required in this stage is T 4 +T 5. The PE utilization is 50% in this stage.

183 162 In the final step Computation of FFT components are done. As the size of the longest up/down is N/2 and the number of stages is log 2 N the total computation time is: [1 + N/2 up/down] log 2 (N) 1 unit of time is chosen for each addition. Here the PE utilization is 100% after [1 + N/2 up/down] log 2 (N) time unit. The above information may be summarized as follows Table 1: FRFT calculation of Time complexity of different stages Step Time Required Time Complexity 1 NT 1 O (N) 2 N (N-1)/2*T 1 + T 2 log 2 N + O (N 2 ) T 3 3 T 4 +T 5 O (1) 4 [1 + N/2 up/down] log 2 (N) O (N log 2 N) 6. Re-configurable FRFT Architecture In our proposed architecture we propose a re-configurable processing element, which can be dynamically reconfigured for different stages of FRFT. The processing element is controlled by a control unit, which generates control signal to do the necessary reconfiguration. As the architecture is systolic, each processing element has two inputs H in & V in used for sending necessary reconfiguration signals and two outputs H out & V out [20]. The process of reconfiguration is discussed stepwise Firstly, eigen vector elements V np is systolically transferred through H in and the vector components is stored in the register R1 and R2. R3 will initially store the zero and R1 is bypassed to H out. Secondly, the signal component V np transferred through H in is multiplied with the register value of R1 and then this value is added with the signal component X (n, p-1) transferred through V in and stored to the register R3. Then the value is by passed through V out. In the third stage the signal H in is stored in the register R1. Then this valued is transferred to H out. Next the value of the register R1 and R2 is multiplied and stored in the register R3. The output is the vector Z K [p]. Finally, the value of the signal element H in is stored to the register R2 and the value of the register R3 and R2 is multiplied and stored to the register R4. This finally yields the transform vector X α [k] Fig 4: Reconfigurable Processing Element for computation of MA- CDFRFT 7. Conclusions In this paper we discussed a re-configurable architecture that computes MA-CDFRFT transforms in four stages. The algorithm developed has a complexity of order (N 2 log N). We discuss some other features that could make the architecture more versatile. Firstly the architecture could be made more Fault Tolerant so that a failed PE is disabled. The rest of the system continues to function as usual. Secondly since there is multiple PEs, there can be multiple simultaneous users of the system, each executing a different task. Thirdly we could use systolic rings to improve inter PE communication instead of systolic array. This architecture could be extended to computation of other image processing algorithms like DCT, DWT, FFT [15] and DFT. This could lead to the development of a generalized transform processor in which a single PE, upon the effect of a control signal, could compute various transforms.

184 163 Acknowledgments We are indebted to Prof. Amitabha Sinha, Director, School of Information Technology, West Bengal University of Technology, Kolkata for his help in this work. References [1]Juan Gaspar Vargas-Rubio, The Central Discrete Fractional Fourier Transform, properties, Computation and application to linear chirp signals, Ph. D thesis, The Univ. of New Mexico, Albuquerque, New Mexico, Dec [2] Machiraju Vijay and C. Siva Ram Murthy, Real-Time Simulation of Dynamic Systems on Systolic Arrays, IEEE Transactions on Industrial Electronics, Vol. 45, No.2, April 1998, pp [3] Rajiv Saxena and Kulbir Singh, Fractional Fourier transform: A novel tool for signal processing J. Indian Inst. Sci., Jan. Feb. 2005, 85, pp [4] T Willey, T S Durrani and R Chapman An FFT Systolic Processor And Its Applications. [5] Griselda Saldaña and Miguel Arias-Estrada, FPGA-Based Customizable Systolic Architecture for Image Processing Applications, Proceedings of the 2005 International Conference on Reconfigurable Computing and FPGAs (ReConFig 2005), IEEE Computer Society. [6] S. Barua, J. E. Carletta, K. A. Kotteri and A. E. Bell, An Efficient Architecture for Lifting-based Two-Dimensional Discrete Wavelet Transforms GLSVLSI 04, April 26 28, 2004, Boston, Massachusetts, USA. [7] Rafael C. Gonzalez, Richard E. Woods, Steven L. Eddins Digital Image Processing using MATLAB Pearson Education. [8] Chris Dick Computing the Discrete Fourier Transform on FPGA Based Systolic Arrays. [9] Xilinx, Introduction and overview, Virtex-II Pro Platform FPGAs, March 9th, [10] K.Sapiecha and R.Jarocki, Modular Architecture For High Performance Implementation Of FFT Algorithm, 1986 IEEE. [11] Rami A. AL Na'mneh, W. David Pan and B. Earl Wells, Two Parallel Implementations for One Dimension FFT On Symmetric Multiprocessors, ACMSE '04, April 2-3, 2004, Huntsville, Alabama, USA, pp [12] S.Y. Kung, VLSI Array Processor, Prentice Hall International Inc., ISBN: X, [13] Pavel Sinha, Amitabha Sinha, Dhruba Basu, A Novel Architecture of a Re-configurable Parallel DSP Processor, IEEE Int. Conf Proc. NEWCAS 05, June 19-22, 2005, pp [14] Kai Hwang and Faye A. Briggs, Computer Architecture and Parallel Processing McGraw-Hill, [15] Preston A. Jackson, Cy P. Chan, Jonathan E. Scalera, Charles M. Rader, and M. Michael Vai, MIT Lincoln Laboratory, A Systolic FFT Architecture for Real Time FPGA Systems. [16] Amitava Sinha, Pavel Sinha, Santanu Chatterjee and Dhruba Basu An Efficient Re-Configurable Architecture of Centered Discrete Fractional Fourier Transform Processor. [17] C. Dick, Computing multidimensional DFTs using Xilinx FPGAs, ICSPAT 98, Toronto, Canada, Sept [18] N.I. Cho and S.U. Lee, DCT algorithms for VLSI parallel implementations, IEEE Trans. Acoust., Speech, Signal Processing, vol. 38, Jan. 1990, pp [19] Soumen Mukherjee, Anal Acharya, An Efficient Reconfigurable Architecture for Fractional Fourier Transforms, International Conference in Signals, Systems and Automation (ICSSA-09) held on December, 2009 in CGET, Anand, Universal Publisher, Page 85-88, ISBN 10: , ISBN -13: [20] Anal Acharya Soumen Mukherjee, Designing Fractional Fourier Transforms using Systolic Arrays National Conference on Emerging Trends in Computer Science and Information Technology (ETCSIC) Nashik held from 29-30th January 2010, in K K Wagh Institute of Engineering Education and Research, Page Anal Acharya is currently the Head of the Department of Computer Science in St. Xavier s College, Kolkata, India His present research interests include Distributed Computing, Object Oriented Modeling & Signal Processing Architecture. He has several published papers in National & International conferences and Journals. He has above 10 years of experience in undergraduate and postgraduate teaching & supervised several post graduate dissertations. Soumen Mukherjee is with RCC Institute of Information Technology Kolkata, India His present research interests include Object Oriented Modeling & Signal Processing Architecture and collaborative learning. He has eleven published papers in National & International conferences and Journals. He has supervised several postgraduate dissertations. He is a Life Member of CSI, IETE, ISTE, ISCA and FOSET. He has also served as a co-opted member in the Executive Committee in the IETE Kolkata Center, India.

185 164 Syllables Selection for the Development of Speech Database for Punjabi TTS System Parminder Singh 1 and Gurpreet Singh Lehal 2 1 Dept. of Computer Science & Engineering, Guru Nanak Dev Engineering College, Ludhiana, Punjab , India 2 Dept. of Computer Science, Punjabi University, Patiala, Punjab , India Abstract The Selection of the speech unit and then the number of speech units for the speech database is one of the important and tedious job. Syllables have been reported as good choice of speech unit for speech database of many languages. For this work also, syllables have been selected as the speech unit for the development of the Punjabi speech database. For minimizing the database size, efforts have been made for the selection of the minimal set of syllables covering almost whole Punjabi word set. To accomplish this all Punjabi syllables have been statistically analyzed on the Punjabi corpus having more than 104 million words. This analysis helped to select a relatively smaller syllable set (about first ten thousand syllables (0.86% of total syllables)) of most frequently occurring syllables having cumulative frequency of occurrence (FOO) less than 99.81%, out of total available syllables. Also to improve the efficiency of the text-to-speech (TTS) system; interesting facts about Punjabi syllables have been obtained based on their FOO at the three (starting, middle and end) positions in the words. Keywords: Speech database, Punjabi syllables, Punjabi TTS system. 1. Introduction Speech database is a core component of a concatenative TTS system. The quality of the output synthesized speech signal depends upon how accurately speech database has been prepared [1]. Two main factors affect the quality of speech database. First, the selection of speech unit, and second the number of speech units to be stored in the database. The first factor controls the naturalness of the output speech signal and the second one affects the database size and hence the response time and portability of the TTS system. Since Punjabi is a syllabic language [2], so syllable has been selected as the basic speech unit for this TTS system, which preserves within unit coarticulation effects [3, 4]. Further, for limiting the number of syllables in the database, the frequency of occurrence of all Punjabi syllables in Punjabi corpus has been found. The syllables occurring very less in the corpus have been ignored and are not selected for storage in the database. 1.1 Punjabi language Punjabi is an Indo-Aryan language spoken by more than 100 million people those are inhabitants of the historical Punjab region (in north western India and Pakistan) and in the Diaspora, particularly Britain, Canada, North America, East Africa and Australia. It is written from left to right using the Gurmukhi (an abugida derived from the Laṇḍā script and ultimately descended from Brahmi script) as well as Shahmukhi (a version of the Arabic script) scripts. In Gurmukhi script, which follows the one sound-one symbol principle, the Punjabi language has thirty eight consonants, ten non-nasal vowels and same numbers of nasal vowels (see Fig. 1). ਸ ਹ ਚ ਛ ਜ ਝ ਞ ਤ ਥ ਦ ਧ ਨ ਯ ਰ ਲ ਵ ੜ ਕ ਖ ਗ ਘ ਙ ਟ ਠ ਡ ਢ ਣ ਪ ਫ ਬ ਭ ਮ ਸ਼ ਖ਼ ਗ਼ ਜ਼ ਫ਼ ਲ਼ ਇ ਈ ਏ ਐ ਅ ਆ ਔ ਉ ਊ ਓ ਇ ਈ ਏ ਐ ਅ ਆ ਔ ਉ ਊ 1.2 Punjabi Syllables Non-nasal Vowels Nasal Vowels Fig. 1 Punjabi Consonants and Vowels Consonants Defining a syllable in a language is a complex task. There are many theories available in phonetics and phonology to define a syllable. In phonetics, the syllables are defined based upon the articulation [5]. However in phonological approach, the syllables are defined by the different sequences of the phonemes. So, combination of phonemes

186 165 gives rise to next higher unit called syllable. Further, combination of syllables produces larger units like morphemes and words. So, syllable is a unit of sound which is larger than phoneme and smaller than word. In every language, certain sequences of phonemes and hence syllables are recognized. Using these phonetic sequences and hence structures, all practically possible syllables can be formed those have been discovered so far in ancient and recent literary works. In addition, all theoretically possible syllables can be composed that may or may not yet be used in a language, but valid in the sense that these follow rendering rules for the language at present [6]. A syllable must have a vowel, without vowel, syllable can not exist. In Punjabi seven types of syllables are recognized [2] V, VC, CV, VCC, CVC, CVCC and CCVC (where V and C represents vowel and consonant respectively), which combine in turn to produce words. The occurrence of syllables of last type CCVC is very rare, and has not been considered in the present work. As said above, Punjabi language has thirty eight consonants, ten non-nasal vowels and same numbers of nasal vowels; so, the above said seven syllable types results syllables in Punjabi with non-nasal vowels and the same number of syllables with nasal vowels and thus giving total of syllables in Punjabi. 2. Statistical Analysis For the development of speech database for this Punjabi TTS system, the syllables of first six types (V, VC, CV, VCC, CVC and CVCC) have been selected and the syllables of type CCVC have not been considered due to their less existence [2]. For selecting syllables for the speech database, the total available syllables (nasal and non-nasal) have been statistically analyzed on a Punjabi corpus. For this purpose a carefully selected balanced corpus having total and unique words have been used. Frequency of occurrence of the available syllables in the corpus has been found. It has been observed that certain syllables are having frequency of occurrence zero; and these syllables have been declared as invalid syllables in Punjabi. The table 1 and table 2 shows the result of the statistical analysis and shows the valid non-nasal and nasal syllables (having FOO>0) of each type respectively. Table 1: Statistical analysis result of non-nasal syllables Syllable Type Total No. of Syllables Syllables with Freq. of Occurrence >0 V VC CV VCC CVC CVCC Total = (Syllables) = (Non zero freq. syllables or valid syllables ) Table 2: Statistical analysis result of nasal syllables Syllable Type Total No. of Syllables Syllables with Freq. of Occurrence >0 V VC CV VCC CVC CVCC Total = (Syllables) = 4218 (Non zero freq. syllables or valid syllables) During utterance stress, duration and articulation of the same syllable is different at the starting, middle and end positions of the words [7]. Hence for the better accuracy of the TTS system, the syllables sounds need to be stored separately for these three positions in the speech database. So, FOO at these three positions has been found separately as shown in Table 3 and Table 4 for non-nasal and nasal syllables respectively. Table 3: Frequency of occurrence at starting, middle and end positions of Non-nasal Syllables Freq. of Occurrence at Syllable Type Starting Position Middle Position End Position Total Freq. V VC CV VCC CVC CVCC

187 166 Table 4: Frequency of occurrence at starting, middle and end positions of Nasal Syllables Freq. of Occurrence at Syllable Type Starting Position Middle Position End Position Total Freq. V VC CV VCC CVC CVCC It has been observed that the occurrence of syllables (nonnasal and nasal) at the starting position in the words is more than at middle and ending positions. Results show that there are 18013, 6251 and syllables having FOO , and at the starting, middle and end positions respectively. The analysis reveals that the syllables occurring at middle position of words are very less than at starting and end positions. Fig. 2 shows the plot for number of times the syllables are occurring at staring, middle and end positions in the words of the said Punjabi corpus. Fig. 3 Number of syllables with frequency in the range Fig. 4 shows the plot for number of syllables against the FOO in the range 1000 to Fig. 2 Syllables frequency at starting, middle and end positions It has been observed that there are 6674 syllables with unit FOO. This is mostly because of the words those are borrowed from other languages and are occurring very rare in Punjabi and their one time existence in the corpus has given the unit FOO to its syllables, those are not otherwise occurring in general in the language. There are syllables having their frequency in the range 1 to 100. Fig. 3 shows the plot for number of syllables against the FOO in the range 100 to Fig. 4 Number of syllables with frequency in the range Fig. 5 shows the plot for number of syllables against the FOO in the range to The number of syllables with FOO more than is only 39.

188 167 Fig. 5 Number of syllables with frequency in the range The combined sorted list of nasal and non-nasal syllables results only syllables, out of total , having FOO (total of three positions) more than zero. Out of these, first (appox.) most frequently occurring syllables having cumulative FOO less than 99.81% have been selected for the development of speech database. Only the syllables having FOO less than 8 have been omitted and these will not affect the working of the TTS system because of their very less occurrence. So with these many syllables, the TTS system will be able to cover almost all Punjabi words as well as would be able to generate words borrowed from other languages and names of persons; and hence producing a general TTS system for Punjabi. The word coverage (number of unique words in which a particular syllable is occurring) by the syllables provides important information about the syllables and it has also been found for the unique words of the above said corpus. Results show that non-nasal syllable type CV is having maximum number of word coverage ( unique words) than other syllables. Fig. 6 and Fig. 7 show these results for the non-nasal and nasal syllables respectively. 3. Conclusions Fig. 6 Word coverage by non-nasal syllables Fig. 7 Word coverage by nasal syllables It has been observed from the above results that the statistical analysis of the Punjabi syllables over the Punjabi corpus plays a vital role in selection of syllables for the speech database. Results show that large number of syllables are not occurring even once in the corpus of about 104 million words and are declared as invalid syllables. Also, other large number of syllables are having comparatively very little frequency of occurrence and are ignored for the final selection. Also for improving the quality of speech database the syllables have been analyzed statistically for the three (staring, middle and

189 168 end) positions in the words of the corpus. This statistical analysis helped to select a relatively small syllable set (of about first ten thousand syllables, that are about 0.86% of total syllables) of most frequently occurring syllables having cumulative frequency of occurrence less than 99.81%, out of available syllables. The results of this statistical analysis will also be very much helpful for the implementation of the other syllable based NLP systems. Engine for Punjabi, Devnagri to Urdu and Urdu/Kashmiri to Roman script transliteration systems, Punjabi morphological analyzer, and intelligent predictive Romanized typing utility for Gurmukhi text. His field of expertise is Natural Language Processing. He is team leader for many projects funded from abroad and Govt. of India. References [1] N. Kalyani, and K.V.N. Sunitha, Syllable analysis to build a dictation system in Telugu language, International Journal of Computer Science and Information Security, Vol. 6, No. 3, 2009, pp [2] P. Singh, Sidhantik Bhasha Vigeyan, Patiala (India): Madan Publications, [3] E.V. Raghavendra, S. Desai, B. Yegnanarayana, A.W. Black, and K. Prahallad, Global syllable set for building speech synthesis in Indian languages, in IEEE Workshop on Spoken Language Technology, Goa (India), 2008, pp [4] M.L. Narayana, and A.G. Ramakrishnan, Defining syllables and their stress in Tamil TTS corpus, in Workshop on Image and Signal Processing, IIT Guwahati (India), 2007, pp [5] R.A. Krakow, Physiological organization of syllables: a review, Journal of Phonetics, Vol. 27, 1999, pp [6] R.K. Joshi, K. Shoff, and S.P. Mudur, A phonemic code based scheme for effective processing of Indian Languages, in 23 rd Internationalization and Unicode Conference, Prague, [7] H. Lee, and C. Seong, Experimental phonetic study of the syllable duration of Korean with respect to the positional effect, in The Fourth International Conference on Spoken Language Processing (ICSLP'96), Philadelphia (USA), 1996, pp Parminder Singh is presently working as Associate Professor in Department of Computer Science and Engineering at Guru Nanak Dev Engineering College, Ludhiana (India). He is M.Tech. and B.Tech. in Computer Science and Engineering and is currently pursuing Ph.D. His field of interest is Natural Language Processing (Speech Synthesis). He has guided sixteen M.Tech. thesis and has published about fifteen research papers in international and national journals and conferences. Gurpreet Singh Lehal is Professor in Department of Computer Science and Director of Advance Centre for Technical Development of Punjabi Language Literature and Culture at Punjabi University, Patiala (India). He is a member of IEEE and ACM. He has about sixty publications to his credit in international journals and conferences. He has developed software like Punjabi Word Processor named Akhar, OCR for Gurmukhi, Gurmukhi-Shahmukhi transliteration system, Search

190 169 Personalized access towards a mobile neuroscience EL ALLIOUI Youssouf 1, EL BEQQALI Omar 2 1 University Department Computer Science Department, Faculty of Sciences Dhar el mahraz, University Sidi Mohammed Ben Abdellah, Fez, 30000, Morocco 2 University Department Computer Science Department, Faculty of Sciences Dhar el mahraz, University Sidi Mohammed Ben Abdellah, Fez, 30000, Morocco Abstract A crucial requirement for the context-aware service provisioning is the dynamic retrieval and interaction with local resources, i.e., resource discovery. The high degree of dynamicity and heterogeneity of mobile environments requires to rethink and/or extend traditional discovery solutions to support more intelligent service search and retrieval, personalized to user context conditions. Several research efforts have recently emerged in the field of service discovery that, based on semantic data representation and technologies, allow flexible matching between user requirements and service capabilities in open and dynamic deployment scenarios. Our research work aims at providing suitable answering mechanisms of mobile requests by taking into account user contexts (preferences, profiles, physical location, temporal information ). In recents works, we have built an ontology, called O Neurolog, to capture semantic knowledge a valuable in Neurology domain in order to assist users (doctor, patient, administration ) when querying Neurology knowledge bases in mobile environment. This current paper focuses specifically on answering mechanisms when accessing to neurological data stored on mobile devices. We present some insights in order to overcome the problem of semantical and personalized access by using similarity between Trees formalizing user needs/requests and available neurological resources. Keywords : Neurology; ontology; context-aware; semantic web; query answering; mobile environment. 1. Introduction 1.1 General context In context-aware information provisioning scenarios, it is crucial to enable the dynamic retrieval of available knowledges in the nearby of the user s current point of attachment, while minimizing user involvement in information selection. Data and knowledge discovery in pervasive environments, however, is a complex task as it requires to face several technical challenges at the state of the art, such as user/device mobility, variations (possibly unpredictable) in service availability and environment conditions, and terminal heterogeneity. Users might need to discover knowledges whose names and specific implementation attributes cannot be known in advance, while data providers need to use several and different terms or keywords and whose technical capabilities and conditions at interaction time might be mostly unpredictable beforehand. In medical domain, there is a great need for using mobile devices to access and retrieve neurological data concerning a patient by physicians or interested organisms (insurance, emergency ). Neurological information is available via web pages, stored in ftp sites or relational databases, and textually described in publications. Neurology (from Greek, neuron, "nerve"; and logia, "study") is a medical specialty dealing with disorders of the nervous system. Specifically, it deals with the diagnosis and treatment of all categories of disease involving the central, peripheral, and autonomic nervous systems, including their coverings, blood vessels, and all effectors tissue, such as muscle.[1] The corresponding surgical specialty is neurosurgery. A neurologist is a physician who specializes in neurology, and is trained to investigate, or diagnose and treat neurological disorders. Pediatric neurologists treat neurological disease in children. Neurologists may also be involved in clinical research, clinical trials, as well as basic research and translational research.[2] However, mobile search engines are unable to answer questions about this massive neurological knowledge base other than identifying resources that contain some subset of the specified attributes. The main reason for this limitation is that the representation of biological information on the web is not machine understandable, in the sense that computers cannot interpret words, sentences or diagrams so as to correctly reason about the objects and the relations between them that are implicitly stated in those documents [3]. The primary goal of the semantic web is to add semantics to the current Web, by designing

191 170 ontologies which explicitly describe and relate objects using formal, logic-based representations that a machine can understand and process [4]. This ongoing effort is expected to facilitate data representation, integration and question answering, of critical importance in the life sciences and hospital information system (HIS). Therefore, returned answers scope needs to be filtered according to finer-grained criteria other than administrative or network grouping. All and only those data that are semantically compatible with the user s context should be automatically and transparently made visible to him. The exploitation of user s contextawareness in knowledge discovery helps mobile clients saving time and efforts in information retrieval. On the other hand, the potential of semantic-based discovery has not been fully exploited yet because of various management issues, which seem to be still open. Access terminals usually exhibit relevant differences in resource capabilities, such as display size and resolution, computing power, memory, network bandwidth, and battery. A crucial management crucial issue remains how to provide support for semantic-based discovery to mobile devices with limited capabilities. Semantic support services, e.g., ontology repositories, inference engines and knowledge management tools, typically require a large amount of computational/memory resources that may not fit the properties of mobile devices. In particular, strict limitations exist about the kind of semantic support facilities that can be hosted on resource-constrained devices. For example, executing a reasoning process on board of a resource-limited device, such as a smart phone, might not only consume battery, but more important, it would probably monopolize all available memory resources, thus making the execution of other applications very difficult. 1.2 Runing example Let us suppose a physician who needs to consult a patient s clinical data in order to set a proper treatment for him. If the healthcare act is taking place inside the hospital, the doctor will be allowed to access the Hospital Information System (HIS) and to retrieve all the patient s Electronic Health Records (EHRs). Having enough time and knowledge and depending on the usability of the software system the specialist will rule out all the useless pieces of information and will get the ones he is interested in. In the latter situation, a brief report including those pieces of the patient s clinical data which ought to be considered would be very valuable. The clinical procedure which is going to be carried out would determine which data should be part of this summary. For example, is an another member of the patient s family has the same symptoms yet. If true the physician could propose muscular or hepatic biopsy and so realize lumbar punctures, cerebral scanner or IRM. The patient does not walk, does not stand nor sitting does not his head. He/she did not speak, but knows how to understand, she established contacts with people who are very familiar. We think that by joining context to domain knowledge could help improving the summarization of research results for a mobile user. This activity is called personalization and implies recommendations. 1.3 Paper organization The remaining of this document is structured as follows. Section II presents a recommender System. The vision of Semantic Web in neuroscience domain is presented in Section III. In Section IV, we propose a semantic and personalized strategy based on answering mechanisms from mobile devices. Section V is devoted to giving some steps of building a prototype called NeuroService 1.0. Finally, in Section VI some conclusions and directions for future work are pointed out. 2. what s a recommander system? As shown in Fig. 1, a recommender system can be running either remotely in a server, or locally in a fixed or mobile consumer device. In both scenarios, the personalization tool selects automatically items that match the users preferences and needs, which are previously modeled in their personal profiles. In current approaches, the profiles store items which are (un)appealing to the users, along with their main attributes (named content descriptions) and their ratings (i.e., the user s levels of interest). These ratings can be explicit or implicit. In the first case, users are required to explicitly specify their preferences for any particular item, usually by indicating a value in a continuous range (e.g., [1, 1]). Negative values commonly mean disliking, while positive values express liking.

192 171 originates most of the weaknesses identified in their personalization strategies, as we will see in Section 2. Personalization strategies for the items recommendation are shown in fig.2 Figure 1. An overview of a recommender system As explicit ratings impose additional efforts on users, recommender systems can also infer information about their interests from their behavior in a much less obtrusive way. Typical examples of implicit ratings are purchase data, reading time of Usenet news, and browsing behavior [5]. Owing to the difficulty of acquiring explicit ratings, some providers of recommendation services adopt hybrid approaches: they compute recommendations based on explicit ratings whenever possible; in case of unavailability, inferred implicit ratings are used instead. Once the user s preferences have been modeled, the recommender system elaborates suggestions by resorting to different personalization strategies. After a recommendation is received, the user can provide information about its accuracy in an explicit or implicit way, analogously to what we have just explained before. As shown in Fig. 1, that information (commonly named relevance feedback) allows the recommender system to update the user s profile, and therefore, to adapt the offered suggestions to the changes in his/her personal preferences. In order to decide whether a given item must be suggested to a user, some personalization strategies compare its attributes with those defined in his/her profile, whereas other techniques miss these content descriptions and only consider the ratings assigned by the users to their preferences. All of these strategies have a common drawback in the fact that the recommendations are made by syntactic mechanisms, which disregard a huge amount of knowledge that may be hidden behind the semantics (i.e., meaning) of both the items content descriptions and the user s preferences. Such a limitation reduces the quality of the suggestions offered by the current recommender systems and, besides, Figure 2. Answering mecanisms 3. Vision of the symantic web and neuroscience 3.1 Neuroscience Neuroscience is in need of a new informatics framework that enables semantic integration of diverse data sources [6]. Experimental data is collected across different scales, from cell to tissue to organ, using a wide variety of experimental procedures taken from diverse disciplines. Unfortunately the information systems holding these data do not link related data among them, preventing effective research that could combine the data to achieve new insights. Integrative neuroscience research is key to providing a better understanding of many neurological diseases such as Alzheimer s disease and Parkinson s disease, and could potentially lead to a better prevention, diagnosis and treatment of such diseases. Identify applicable sponsor/s here. If no sponsors, delete this text box. (sponsors)

193 172 The field of computer consultation has passed through three historical phases. In the first, attempts were made to improve on human diagnostic performance by rejecting the methods used by clinicians and substituting various statistical techniques [7]. Statistical methods proved to be accurate for small diagnostic domains, but impractical for application to realworld problems [8]. In the second phase, it was recognized that human problem solving methods were deceptively powerful [9] and attempts were made to capture diagnostic logics as fixed decision protocols [10]. Although these met with success in some areas, it was recognized that such protocols suffered from inflexibility [11]. At present, efforts are directed towards building systems which incorporate expert problem solving strategies, but which retain flexibility - artificial intelligence systems [12]. Neurology is a medical specialty concerned with the diagnosis and treatment of all categories of disease involving the central, peripheral, and autonomic nervous systems, including their coverings, blood vessels, and all effector tissue, such as muscle. [13] Neurological disorders are disorders that can affect the central nervous system (brain and spinal cord), the peripheral nervous system, or the autonomic nervous system. Conditions can include but are not limited to: - Brain injury, spinal cord and peripheral nerves - Cerebral palsy - Cerebrovascular disease, such as transient ischemic attack and stroke. - Epilepsy - Headache disorders such as migraine, cluster headache and tension headache. - Infections of the brain (encephalitis), brain meninges (meningitis), spinal cord (myelitis) Semantic web The Semantic Web, a maturing set of technologies and standards backed by the World Wide Web consortium [14], offers technical guidance specifically in the area of aggregating and integrating diverse information resources. These Semantic Web technologies can be used to integrate neuroscience knowledge and to make such integrated knowledge more easily accessible to researchers. The foundational technologies of the Semantic Web Resource Description Framework (RDF [15]), Web Ontology Language (OWL [16]), the SPARQL Protocol and RDF Query Language (SPARQL) are widely implemented and are backed by a large community of users and developers. The chief advantages of Semantic Web technologies include (1) the widely supported standards backed by the World Wide Web consortium, (2) the ability to make use of the well-established inference mechanisms of description logics, and (3) the availability of a wide range of software tools. 3.3 Ontology Ontologies are defined as formal, explicit specifications of a shared conceptualization [17], encode machineinterpretable descriptions of the concepts and the relations in a domain using abstractions as class, role or instance, which are qualified using logical axioms. Properties and semantics of ontology constructs are determined by Description Logics (DLs) [18], a family of logics for representing structured knowledge which have proved to be very useful as ontology languages. Ontologies have become the cornerstone in the Semantic Web due to two reasons. On the one hand, as these conceptualizations represent formally a specific domain, they enable inference processes to discover new knowledge from the formalized information. On the other hand, ontologies facilitate automated knowledge sharing, by allowing easy reuse between users and software agents. The last feature was first promoted by standards like RDF [19] and RDFS [20], which added a formal semantics to the purely syntactic specifications provided in XML. Next, DAML (DAML: The DARPA Agent Markup Language, 2000) and OIL [21] arose, which have been finally fused and standardized by W3C as OWL [22]. Nowadays, OWL is the most expressive language in which three sublevels have been defined (Lite, DL and Full). In this regard, the language used to implement the ontology required in our reasoning approach depends on the knowledge and expressive necessities of each application domain and each recommender system. 4. Answering mechanisms from mobile devices 4.1 Formalizing domain knowledge and modeling user preferences In the field of Semantic Web, an ontology is a formal specification of a conceptualization [23], that is, an abstract and simplified view of the world that we wish to represent, described in a language that is equipped with a formal semantics. An ontology characterizes that semantics in terms of concepts and their relationships, represented by classes and properties, respectively. Both entities are

194 173 hierarchically organized in the conceptualization, which is populated by including specific instances of both classes and properties. For example, in the context of a recommender system, instances of classes represent the available items and their attributes, whereas instances of properties link the items and attributes to each other. All these semantic knowledges are formalized and specified in OWL ontologies with Protégé2000 editor. In order to reason about the user s preferences, our approach needs a formal representation which includes semantic descriptions of items that have been interesting or unappealing to him/her (named positive and negative preferences, respectively). These descriptions allow the recommender system to learn new knowledge about the user s interests, which will be exploited during the reasoning-based recommendation process. Figure 4. Our ontology-based approach of modeling user in Oprofile. 4.2 Basical principes of mobile answering strategy In order to fight the aforementioned limitations, our personalization approach defines a metric that compares the user s preferences and the available items in a flexible way: instead of using syntactic techniques, we reason about the semantics of the compared items. For that purpose, we take advantage of the inference mechanisms involving semantic descriptions developed in the Semantic Web. The use of semantic information in recommender systems has been already proposed in various systems. In the simplest proposals, the semantic descriptions serve to provide the users with additional information about the items they have rated. Figure 3. A brief excerpt from an ontology about the Neurology domain. Our approach models the user s preferences by reusing the knowledge formalized in the domain ontology. As the available items, their attributes and the hierarchical categories are already defined in the conceptualization, our user s models only maintain references to the instances that identify his/her preferences in the ontology. In previous work [24], we have yield a semantic model to capture the neurological domain knowledge and the user preferences, as : - Domain knowledge about the neurology sciences, named O Neurolog - Profile knowledge grouped by the preferences and needs of the user (patient, Doctor, Lawyer, Banc ), named O Profile. The approach we propose in this paper fights the limitations of the traditional syntactic strategies by taking advantage of the experience gained in the Semantic Web field. According to the guidelines established by Berners- Lee et al. (2001). Semantic Web is based on describing Web resources by semantic annotations (metadata), formalizing these annotations in ontology, and applying reasoning processes aimed at discovering new knowledge. Specifically, our approach improves the personalization capabilities of the current recommender systems by resorting to a strategy based on semantic reasoning. For that purpose, we lean on a domain ontology in which the semantic descriptions of the available items (e.g., drugs, diseases...) are formalized. Instead of employing the traditional syntactic approaches, our reasoning based strategy discovers semantic relationships between the users preferences and the items available in the domain

195 174 ontology. These relationships provide the system with extra knowledge about the user s interests, thus favoring more accurate personalization processes. The basical principles of mobile answering mechanisms are: - Formalize O profile and O neurological as knowledge and data Trees: Abstraction Phase - Send user neurological needs as keywords : Formulation Phase - Compute similarity between keyword and O neurolog Ontology concepts/properties : Matching Phase - Retrieve from all neurological sources items relevant with concepts/properties more similar : Collaborative answering Phase - Filter and Reduce results according to user preferences, as specified in O profile : Pruning Phase - Display those results by using Visualization methods according to technical characteristics of mobile devices : Presentation Phase Neurological resources are available on diverse servers as XML Documents according to autonomous, heterogenous, and distributed XML Schemas. Mobile devices are used by users (patients, doctors, medical partners ) in order to access to these data and knowledges. In this context, querying heterogeneous collections of data-centric XML documents requires a combination of database languages and concepts used in information retrieval, in particular similarity search and ranking and their adaptation to mobile context. In order to improve these principles, our current work focus is based on determining the degree of similarity, called DoS between a keyword and a Concept or a Property according to terminological, structural and semantical criteria. So we could exploit many researchs done by RI Community in Tree Matching or Complex Object Mapping. In particular, Tree Embedding algorithm could be useful for this kind of problem. Figure 5. ApproXQL Query TheMapping of an approxql query to a Query Tree could be done as follows : Figure 6. Query Tree. Authors have shown how to interpret both the data and the query as trees according to several scenarios. With this interpretation, the problem of answering a query can be mapped to the problem of embedding a query tree (User needs) in the data tree (available Neurological resources). Figure 7. Unordered inclusion of a query tree in a data tree. The goal is to approximately embed the query tree into the data tree such that the labels and the ancestorship of the nodes are preserved: We present here an improvement of this problem, as proposed by T. Schielder in the paper [25]. In fact, User Query could be formalized as a Query Tree. Let us give an example from a Query b expressed in ApproXQL [25] :

196 175 systems may then interact with the Web service in a manner prescribed by its definition, using XML based messages conveyed by Internet protocols. The Web service architecture defined by the W3C enables application to application communication over the Internet. Web services allow access to software components through standard Web technologies, regardless of platforms, implementation languages, etc. In term of the Internet reference model, the Web service layer could be placed between the Transport and Application Layer. The Web service layer is based on several standard Internet protocols, whereby the protocols WSDL, SOAP, and typically HTTP as depicted in Fig. 8 should be supported by all Web service implementations for interoperability. Figure 8. Approximate Tree Embedding Algorithm 5. Realizing a prototype : NeuroService The technical architecture : web services and mobility Web Service architecture. A Web service is a software system identified by a URI, whose public interfaces and bindings are defined and described using XML [26]. The definition of a Web service can be exported to a file, published to a lookup service, and discovered by other software systems. These The HTTP protocol that builds the first layer of the interoperable part of the protocol stack is, because of its ubiquity, the de facto transport protocol for Web services. But any other transport protocols such as SMTP, MIME, and FTP for public domains as well as CORBA and Message Queuing protocols for private domains could be used instead. The XML-based SOAP forms the next layer. SOAP provides XML-based messaging. In combination with HTTP, XML function calls can be sent as payload of HTTP POST. Because of the extensibility of SOAP, one can define customized messages using SOAP headers. The highest interoperable layer is the XML-based Web Services Description Language (WSDL). A WSDL document serves as a contract to be followed by Web service clients. It defines the public interfaces and mechanisms of Web service interactions. 5.2 mobile techcnologies and languages: a state of art J2ME is a wireless development platform based on Java technology. It is targeted at mobile devices with embedded nature and limited resources. J2ME provides the ability of servers to accept a new set of clients: cell phones, twoway pagers, and palmtops. These devices can be programmed using the Mobile Information Device Profile (MIDP), a set of Java APIs which, together with the CLDC provide a complete Java runtime environment [27]. The main aim behind J2ME is to inherit the powerful features of the Java programming language by designing a light-weight virtual machine (KVM) [28] capable of providing a secure and clean execution environment on resource constrained mobile devices.

197 176 The Android platform delivers a complete set of software for mobile devices: an operating system, middleware, and key mobile applications [29]. Windows Mobile and Apple s iphone provide a richer, simplified development environment for mobile applications. However, unlike Android, they re built on proprietary operating systems that often prioritize native applications over those created by third parties and restrict communication between applications and native data. Android offers new possibilities for mobile applications by offering an open development environment built on an open source Linux kernel. As Fig. 9 illustrates, the Open Mobile Alliance (OHA) [30] Google support the Android platform and hope to reach the goal of ensuring global mobile services that operate across devices, geographies, service providers, operators, and networks. Figure 9. Web Service architecture. The Android platform has recently been ported into mobile devices, such as notebooks, PDAs, and automotive systems. Android software stack consists of a Linux kernel, a collection of Android libraries, an application framework that manages Android applications in runtime, and native or third-party applications in the application layer. 6. Conclusion and perspectives In this paper, we have presented a personalization approach that soothes unresolved limitations of traditional syntactic recommendation strategies by applying semantic reasoning techniques. In fact we present principal insights of an approach to find approximate answers to formal user queries. We reduce the problem of answering queries against XML document collections to the well-known unordered tree inclusion problem. We will extend this problem to an optimization problem by applying a cost model to the embeddings. Thereby we are able to determine how close parts of the XML document match a user query. We present an efficient algorithm that finds all approximate matches and ranks them according to their similarity to the query. To this aim, we take advantage of the knowledge represented in the domain ontology and the semantic relationships that can be inferred from it. Let us recall that we have capitalized semantic knowledges on neurological domain and user preferences. Further, We shall exploit these semantical knowledges to personalized answers to be returned to users when asking mobile devices. Instead of offering items with the same attributes as those defined in the user s profile, our reasoning-based approach suggests items semantically related to his/her preferences, thus diversifying the recommendations. These semantic relationships provide additional knowledge about the user s interests and, therefore, favor more accurate personalization processes. Secondly, the collaborative phase of our strategy allows to select a user s neighbors even when the data about their preferences are very sparse. References [1] urology pdf. Last see 11/04/2010. [2] Last see 11/04/2010. [3] Robu I, Robu V, Thirion B. An introduction to the Semantic Web for health sciences librarians. J Med Libr Assoc 2006;94(2): [4] Berners-Lee T, Hall W, Hendler J, Shadbolt N, Weitzner DJ. Computer science. Creating a science of the Web. Science 2006;313(5788): [5] Montaner, M., López, B., de La Rosa, J.L., A taxonomy of recommender agents on the Internet. Artificial Intelligence Review 19 (4), [6] Martone ME, Gupta A, Ellisman MH. E-neuroscience: challenges and triumphs in integrating distributed data from molecules to brains. Nature Neuroscience 2004;7(5): [7] [accessed ]. [8] [accessed ]. [9] [accessed ]. [10] [accessed ]. [11] [accessed ]. [12] [accessed ]. [13] Marshall MS, Prud hommeaux E. A prototype knowledge base for the life sciences, W3C interest group note. Web publication: [14] [accessed ]. [15] [accessed ]. [16] [accessed ].

198 177 [17] Studer, R., Benjamins, V. R., & Fensel, D. (1998). Knowledge engineering: Principles and methods. Data and Knowledge Engineering, 25(1 2), [18] Baader, F., Calvanese, D., McGuinness, D., Nardi, D., & Patel- Schneider, P. F. (2003). The Description Logic handbook: Theory, implementation, and applications. Cambridge University Press. Semantic'Discoveries [19] Beckett, D., RDF syntax specification. < [20] Brickley, D., Guha, R., RDF vocabulary description language 1.0: RDF Schema. < [21] Fensel, D., van Harmelen, F., Horrocks, I., McGuinness, D., Patel- Schneider, P., OIL: An ontology infrastructure for the Semantic Web. IEEE Intelligent Systems 16 (2), [22] McGuinness, D., van Harmelen, F., OWL Web ontology language overview. W3C Recommendation. [23] Berners-Lee, T., Hendler, J., Lassila, O., The semantic Web: a new form of Web content that is meaningful to computers will unleash a revolution of new possibilities. The Scientific American 279 (5), [24] Y. EL ALLIOUI, O. EL BEQQALI, O Neurolog Building an Ontology for Neurology in a Mobile Environment SITA 2010 : Sixth International Conference on Intelligent Systems ENSIAS, Rabat, 04 and 05 may [25] Torsten Schlieder and Felix Nauman, Approximate Tree Embedding for Querying XML Data [26] W3C, Web services Activity, [27] G. Lawton, Moving java into mobile phones, IEEE Computer [28] Sun Microsystems, Inc., Javae 2 Platform Micro Edition (J2MEe) Technology for CreatingMobile Devices, [29] Android Platform Official Site, < [30] Open Mobile Alliance Official Site, <

199 178 High Performance Direct Torque Control of Induction Motor Drives Using Space Vector Modulation S Allirani 1 and V Jagannathan 2 1 Department of EEE, Sri Ramakrishna Engineering College Coimbatore, Tamilnadu , India 2 Department of EEE, Coimbatore Institute of Technology Coimbatore, Tamilnadu , India Abstract This paper presents a simple approach to design and implement Direct Torque Control technique for voltage source inverter fed induction motor drives. The direct torque control is one of the excellent strategies available for torque control of induction machine. It is considered as an alternative to field oriented control technique. The Direct Torque Control scheme is characterized by the absence of PI regulators, co-ordinate transformations, current regulators and pulse width modulated signal generators. Direct Torque Control allows a good torque control in steady state and transient operating conditions. The direct torque control technique based on space vector modulation and switching table has been developed and presented in this paper. Keywords: Direct Torque Control (DTC), Space Vector Modulation (SVM), Induction Motor (IM). 1. Introduction The induction motor is well known as the work horse of industry. It is estimated that induction motors are used in seventy to eighty percent of all industrial drive applications due to their simple mechanical construction, reliability, ruggedness, low cost and low maintenance requirement compared to other types of motors. Also it operates at essentially constant speed. These advantages are however suppressed from control point of view. When using an induction motor in industrial drives with high performance demands, the induction motors are non linear high order systems of considerable complexity [1]. The advancement of power electronics had made it possible to vary the frequency of the voltage or current relatively easy using various control techniques and thus has extended the use of induction motor in variable speed drive applications. Induction motor control methods can be broadly classified into scalar control and vector control. Scalar control based on relationships valid in steady state conditions. Only magnitude and frequency (angular speed) of voltage, current, and flux linkage space vectors are controlled. Thus, the scalar control does not act on space vector position during transients. Vector control is based on relations valid for dynamic states, not only magnitude and frequency (angular speed) but also instantaneous positions of voltage, current, and flux space vectors are controlled. Thus, the vector control acts on the positions of the space vectors and provides their correct orientation both in steady state and during transients. In scalar control, V/F control is the important control technique, it is the most widespread, reaching approximately 90% of the industrial applications [2]. It acts imposing a constant relation between voltage and frequency. The structure is very simple and it is normally used without speed feedback, hence this control does not achieve a good accuracy in both speed and torque responses mainly due to the fact that the stator flux and the torque are not directly controlled. The speed can be 2% (except in a very low speed) and the dynamic response can be approximately around 50ms. The vector control is most popular method; known as field-oriented control (FOC) gives the induction motor high performance. In the vector control the motor equations are transformed in a coordinate system that rotates in synchronism with the rotor flux vector. These new coordinates are called field coordinates. In field coordinates under constant rotor flux amplitude, there is a linear relationship between control variables and torque similar to a separately excited dc motor. These coordinate transformations are selected to achieve decoupling and linearization of induction motor equations. Main advantages of vector controllers include good accuracy such as 0.5% regarding the speed and 2% regarding the torque, even in stand still. The main disadvantages are the

200 179 huge computational capability required and the compulsory good identification of the motor parameters [3]. The method proposed to replace the de-coupling control with the bang-bang control, which meets very well with on off operation of the inverter semiconductor power devices is referred to as direct torque control (DTC) [5]. The DTC also exploits the vector relationships but replaces the coordinate transformation concept of standard vector control with a form of bang - bang action, dispensing from pulse width modulation (PWM) current control [6]. In DTC the bang-bang or hysteresis controllers impose the time duration of the active voltage vectors, moving stator flux along the reference trajectory. The key features of DTC compared to standard vector control includes No current loops so current not directly regulated Coordinate transformations not required No separate voltage pulse width modulator Stator flux vector and torque estimation required. Recent advancements in DTC systems include the use of unified flux control scheme [11], stator flux vector control in field weakening region [14], torque ripple minimization techniques [12], space vector modulation (SVM) technique [7], neuro - fuzzy [15], FPGA [13]. The purpose of this paper is to review the simulation of DTC technique using MATLAB/SIMULINK model. 2. Principle of Direct Torque Control DTC provides very quick response with simple control structure and hence this technique is gaining popularity in industries [8], [10]. In DTC, stator flux and torque are directly controlled by selecting the appropriate inverter state. The stator currents and voltages are indirectly controlled hence no current feedback loops are required. Nearly sinusoidal stator fluxes and stator currents enable high dynamic performance even at standstill [4]. The generic DTC scheme for a Voltage source PWM inverter-fed IM drive is shown in Fig.1. The scheme includes two hysteresis controllers. The stator flux controller imposes the time duration of the active voltage vectors, which move the stator flux along the reference trajectory, and the torque controller determinates the time duration of the zero voltage vectors which keep the motor torque in the predefined hysteresis tolerance band. At every sampling time the voltage vector selection block chooses the inverter switching state (S A, S B, S C ) which reduces the instantaneous flux and torque errors. 3. Basic Switching Table and Selection of Voltage Vectors The basic idea of the switching table DTC concept is shown in Fig. 1. The command stator flux Ψ sref, and torque T eref values are compared with the actual Ψ s and T e values in hysteresis flux and torque controllers, respectively. The flux controller is a two-level comparator while the torque controller is a three level comparator. The digitized output signals of the flux controller are defined as in equations (1) and (2) 1, for H (1) serr serr s sref 1, for H (2) s sref

201 180 Fig. 2 Inverter voltage vectors and stator flux switching sector And those of the torque controller as in equations (3), (4), (5), T eerr 1, fort e T eref H (3) T eerr 0, fort e T eref (4) Teerr 1, fort e Teref H (5) where 2H Ψ is the flux tolerance band and 2H m is the torque tolerance band. The digitized variables Ψ serr, T eerr and the stator flux section (sector) N, obtained from the angular position α = arctg (Ψ sβ / Ψ sα ) (6) create a digital word where, - 30 o < α (1) < 30 o Fig. 1 Basic scheme of PWM inverter fed induction motor with DTC 30 o < α (2) < 90 o 90 o < α (3) < 150 o 150 o < α (4) < 210 o 210 o < α (5) < 270 o 270 o < α (6) < 330 o On the basis of torque and flux hysteresis status and the stator flux switching sector, which is denoted by α, DTC algorithm selects the inverter voltage vector from the Table1. The outputs of the switching table are the settings for the switching devices of the inverter. Fig.2 shows the relation of inverter voltage vector and stator flux switching sectors. Six active switching vectors V1, V2, V3, V4, V5, V6 and two zero switching vectors V0 and V7 determine the switching sequence of the inverter. Depending on inverter switching pulses, PWM is achieved and hence stator voltages and currents are controlled [3]. Therefore to obtain a good dynamic performance, an appropriate inverter voltage vectors V i (i=1 to 6) has to be selected. Ψ serr 1 0 Table 1: Switching table of Inverter Voltage Vectors T eerr α(1) α(2) α(3) α(4) α(5) sect 1 sect 2 sect 3 sect4 sect5 α(6) sect 6 1 V2 V3 V4 V5 V6 V1 0 V7 V0 V7 V0 V7 V0-1 V6 V1 V2 V3 V4 V5 1 V3 V4 V5 V6 V1 V2 0 V0 V7 V0 V7 V0 V7-1 V5 V6 V1 V2 V3 V4

202 Stator Flux Control By selecting the appropriate inverter output voltage V i (i=1-6), the stator flux Ψ s rotates at the desired frequency ω s inside a specified band. If the stator ohmic drops are neglected, the stator voltage impresses directly the stator flux in accordance with the equations (7) and (8). d s Vs (7) dt d V dt (8) s s Therefore the variation of the stator flux space vector due to the application of the stator voltage vector V s during a time interval of Δt can be approximated as in equation (9). t (9) s V s 3.2. Torque Control 3 P Lm Te sin ' s r (10) 2 2 L s The electromagnetic torque given by equation (10) is a sinusoidal function of γ, the angle between Ψs and Ψ r as shown in Fig.3. The variation of stator flux vector will produce a variation in the developed torque because of the variation of the angle γ between the two vectors as in equation (11). 3 P Lm T e ( ) sin ' s s r (11) 2 2 L s In accordance with the Fig. 1, the flux linkage and torque errors are restricted within its respective hysteresis bands. It can be proved that the flux hysteresis band affects the stator-current distortion in terms of low order harmonics and the torque hysteresis band affects the switching frequency. The DTC requires the flux and torque estimations, which can be performed as proposed in this model, by means of two different phase currents and the state of the inverter. The flux and torque estimations can be performed by means of other estimators using other magnitudes such as two stator currents and the mechanical speed, or two stator currents again and the shaft position [3]. Fig. 3 Stator flux and rotor flux space vectors 4. Simulation Results A direct torque control algorithm of Induction motor drive has been modelled and simulated using Matlab/Simulink simulation package [8], [9]. The simulink model of the three phase induction motor rated 10 HP, 415 V, 1440 rpm is shown in Fig.4. The motor is fed from an IGBT PWM inverter. The MATLAB / SIMULINK model for switching logic is developed and shown in Fig. 5. The transient performance of the developed DTC model has been tested by applying a variable load torque command on the mechanical dynamics. The flux reference is maintained at 0.9 Wb. Figure 6 shows the results obtained. Figure 6 (a) and Fig. 6 (b) shows the electromagnetic torque and Fig 6 (c) shows the rotor speed of the machine. This demonstrates that the developed DTC achieved high dynamic performance in speed response to changes in demand torque. However, there is some performance degradation with torque overshoot in the torque transient owing to the hysteresis controllers used. Figure 6 (d) shows the d axis stator current. Figure 6 (e) shows the stator flux magnitude, emphasizing the decoupled action of torque and flux control. It is observed that the variation of motor torque does not influence fluxes. Figure 6(f) shows the stator current magnitude.

203 182 Fig. 4 MATLAB model for DTC scheme Fig. 5 Simulink model for switching logic of inverter

204 Estimated Torque Torque in Nm Time (secs) Fig.6 (a) Electromagnetic torque 140 Estimated Torque Torque in Nm Time (secs) Fig. 6(b) Initial starting transients of estimated torque

205 Rotor Speed 150 Speed (rps) Time (secs) Fig. 6 (c) Rotor speed 150 d-axis current 100 Current (Amps) Time (secs) Fig. 6 (d) d axis stator current Stator Flux in dq axis Stator flux in d-axis Stator flux in q-axis 0.5 Flux (wb) Time (secs) Fig. 6 (e) Stator flux magnitude

206 Iabc 100 C u rre n t (A m p s ) Time (secs) Fig, 6 (f) Stator current 5. Conclusions The work carried out in this paper is aimed and focused to develop a Simulink model of direct torque control of induction motor drive. The DTC technique allows the independent and decoupled control of torque and stator flux. In order to show the effectiveness of the model, a numerical simulation has been performed on a 10 HP induction machine fed by an IGBT PWM inverter. The feasibility and the validity of the developed DTC model, based on SVM and switching table technique, have been proved by simulation results obtained in the torque control mode. DTC represents a viable alternative to FOC, being also a general philosophy for controlling the ac drives in both motor and generator mode of operation. From a general perspective, FOC requires an accurate estimation of the rotor flux vector. However, when an accurate estimation of the motor flux is available, there is no need to set up a current control loop and DTC is the natural solution. The main features of DTC can be summarized as follows. DTC operates with closed torque and flux loops but without current controllers. DTC needs stator flux and torque estimation and, therefore, is not sensitive to rotor parameters. DTC is inherently a motion-sensor less control method. DTC has a simple and robust control structure; however, the performance of DTC strongly depends on the quality of the estimation of the actual stator flux and torque. References [1] B. K. Bose, Modern Power Electronics and AC Drives. Englewood Cliffs, NJ: Prentice-Hall, [2] J.W.Finch, and D.Giaouris, Controlled AC Electrical drives, IEEE Transactions on Industrial Electronics, Vol. 55, No. 2., 2008, pp [3] G.S. Buja, and P.K.Marian, Direct Torque control of PWM Inverter-Fed AC Motors A Survey, IEEE Transactions on Industrial Electronics, Vol. 1, No. 4, 2004, pp [4] M. P. Kazmierkowski, and A. Kasprowicz, Improved direct torque and flux vector control of PWM inverter-fed induction motor drives, IEEE Transactions on Industrial Electronics, Vol. 42, No.4, 1995, pp [5] I. Takahashi, and T.Noguchi, A new quick response and high efficiency control strategy of an induction motor, IEEE Transactions on Industry Applications, Vol.1A-22, No.5, 1986, pp [6] G. Buja, D. Casadei, and G. Serra, DTC- Based strategies for induction motor drives, in Proc. IEEE IECON 97, 1997, pp [7] T. G. Habetler, F. Profumo, M. Pastorelli, and L. M. Tolbert, Direct Torque control of induction motor using space vector modulation, IEEE Trans. on Industry Applications, Vol. 28, 1992, pp [8] H.F. Abdul Wahab, and H. Sanusi, Simullink Model of Direct Torque Control of Induction Machine, American J. of Applied Sciences, Vol.5, No. 8, 2008, pp [9] [10] [11] J.H. Ryu, K. W. Lee, and J.S. Lee, A unified flux and torque control method for DTC based induction motor drives, IEEE Trans. on Power Electronics, Vol. 21, 2006, pp [12] N. R. N. Idris, and A.H.M. Yatim, Direct Torque control of Induction machines with constant switching frequency

207 186 and reduced torque ripple, IEEE Trans. on Industrial Electronics, Vol. 51, 2004, pp [13] S.K. Sahoo, G.T.R. Das, and V.Subrahmanyam, VLSI design approach to high - performance direct torque control of induction motor drives, World J. of Modelling and Simulation, England, Vol.4, No.4, 2008, pp [14] M. Mengoni, L. Zarri, A. Tani, G. Serra, and D. Casadei, Stator flux vector control of Induction Motor drive in the field weakening region, IEEE Trans. on Power Electronics, Vol. 23, 2008, pp [15] P. Z.Grabowski, M. P. Kazmierkowski, B.K. Bose, and F.Blaabjerg, A simple Direct Torque and Neuro Fuzzy control of PWM inverter fed induction motor drive, IEEE Trans. on Industrial Electronics, Vol.47, 2000, pp Dr.V.Jagannathan received his B.E. Electrical Engineering and M.Sc. Engineering from Coimbatore Institute of Technology (CIT), Coimbatore, India in 1965 and 1971 respectively. He completed his Ph.D. in Electrical Engineering from IIT, Kharagpur in He is currently Professor and head of Electrical and Electronics Engineering department in CIT. He has total teaching experience of 40 years. He has published three books in basic electrical and power electronics and 17 papers in leading national and international conferences. Prof.S.Allirani received her B.E. Electrical & Electronics Engineering and M.E. Electrical Machines from Coimbatore Institute of Technology (CIT) in 1994 and PSG College of Technology in 2004 respectively. She is a research scholar of Anna University of Technology, Coimbatore in Electrical Engineering from She is currently Assistant Professor, Electrical and Electronics Engineering department in Sri Ramakrishna Engineering College, Coimbatore. She has total teaching experience of 14 years. She has published 4 papers in national and international conferences.

208 187 A new Morphological Approach for Noise Removal cum Edge Detection M Rama Bai 1, Dr V Venkata Krishna 2 and J SreeDevi 3 1 Dept. of CSE, JNTU (H), M.G.I.T Hyderabad, Andhra Pradesh, India 2 Dept. of CSE, JNTU (K), C.I.E.T Rajahmundry, Andhra Pradesh, India 3 Dept. of CSE, JNTU (H), M.G.I.T Hyderabad, Andhra Pradesh, India Abstract Edge detection is an important aspect in image processing. When a noisy image is presented for edge detection, the noise creates problem in the process of edge detection using conventional methods. One of the disadvantages of the conventional methods is that the noise is not removed automatically. The present paper proposes a novel approach for noise removal cum edge detection for both gray scale and binary images using morphological operations. Two images consisting of noise are processed and the effectiveness of the proposed approach is experimentally demonstrated. The results demonstrate that the proposed filter cum edge detector approach overcomes the deficiency of conventional methods and efficiently removes the noise and detects the edges. Key words: Mathematical morphology, Structuring element, edge detection, morphological residue detector, noise, morphological gradient. 1. Introduction Edges play an important role in image processing hence their detection is very important. The result of the final processed image depends on how effectively the edges have been extracted. The function of edge detection is to identify the boundaries of homogeneous regions in an image based on properties such as intensity and texture. Some early conventional methods for edge detection are Sobel algorithm, Prewitt algorithm and Laplacian of Gaussian operator. But they belong to high pass filtering methods, which are not effective for noisy images because noise and edge belong to the scope of high frequency. In real world applications, images contain object boundaries, object shadows and noise. Therefore, it may be difficult to distinguish the exact edge from noise or trivial geometric features. Many edge detection algorithms have been developed based on computation of the intensity gradient vector, which, in general, is sensitive to noise in the image. Another approach is to study the statistical distribution of intensity values. The idea is to examine the distribution of intensity values of neighbourhood of a given pixel and determine if the pixel is to be classified as an edge. Although there exist some works, less attention has been paid to statistical approaches than the gradient methods in image processing. As the performance of classic edge detectors degrades with noise, edge detection using morphological operators is studied. In this paper, a new morphological approach for noise removal cum edge detection is introduced for both binary and gray scale images. For detecting edges in an image efficiently, first the noise is to be removed. Noise in binary images is of two colours, black and white. The noise in gray scale images manifests itself as light elements on a dark background and as dark elements on the light region. Noise is removed using morphological operations and further morphological operations are applied on this image to extract the edges. 2. Basic Operations of Mathematical Morphology Mathematical morphology is based on set theory which can be used to process and analyse the images. It provides an alternative approach to image processing based on shape concept stemmed from set theory. In mathematical morphology images are treated as sets, and morphological transformations which derived from Minkowski addition and subtraction are defined to extract features in images. The image which will be processed by mathematical morphology theory must be changed into set and represented as matrix. Structuring Elements are used in morphological theory, which are also represented as matrices. Structuring

209 188 element is a characteristic of certain structure and features to measure the shape of an image and is used to carry out other image processing operations. The shape and size of the structuring element (SE) plays crucial role in image processing and is therefore chosen according to the condition of the image and demand of processing. The basic mathematical morphological operations namely dilation, erosion, opening, closing are used for detecting, modifying, manipulating the features present in the image based on their shape. In the following, some basic mathematical morphological operations of gray-scale images are introduced. Let I (x, y) denote a gray-scale two dimensional image, SE denote structuring element. Dilation of a gray-scale image I(x, y), by a gray-scale structuring element SE (a, b) is denoted by (I SE) (x, y) = max{i (x-a, y-b) + SE(a,b)} (1) Erosion of a gray-scale image I (x, y) by a gray-scale structuring element SE(a, b) is denoted by (I SE)(x, y) = min{i (x+a, y+b) - SE(a,b )} (2) Opening and closing of gray-scale image I (x, y) by grayscale structuring element SE(a,b) are denoted respectively by I SE = (I SE) SE (3) I SE = (I SE) SE (4) Erosion basically decreases the gray-scale value of an image by applying shrinking transformation, while dilation increases the gray-scale value of the image by applying expanding transformation. But both of them are sensitive to the image edge whose gray-scale value changes obviously. Erosion filters the inner image while dilation filters the outer image. Opening is erosion followed by dilation and closing is dilation followed by erosion. Opening generally smoothes the contour of an image, breaks narrow gaps. As opposed to opening, closing tends to fuse narrow breaks, eliminates small holes, and fills gaps in the contours. processing. The size and shape of the structuring element decide the final results of detected edges and de-noising in both binary and gray scale images. Three methods are proposed to detect edges in the image. Method 1: The edge of the image is detected by the following process. The edge of an image I which is denoted by E (I) is defined as the difference set of the dilation domain of I and the domain of I. It can be depicted by the following equation {[(I SE) SE] SE} SE {[( I SE) SE] SE} (5) Method 2: The edge of the image can also be detected by the following process. The edge of an image I which is denoted by E (I) is defined as the difference set of the domain of I and the eroded domain of I. It can be depicted by the following equation {[( I SE) SE] SE} - {[(I SE) SE] SE} SE (6) Method 3: The edge of the image can also be detected by the following process. The edge of an image I which is denoted by E (I) is defined as the difference set of the dilated domain of I and the eroded domain of I. It can be depicted as follows {[(I SE) SE] SE} SE - {[(I SE) SE] SE} SE (7) 4. Experimental Results and Conclusions The proposed methods are applied on two images, a binary image with salt and pepper noise as shown in Fig.2 and a gray scale image with salt and pepper noise as shown in Fig.3. The structuring element used in all the methods is a 3x3 window as in Fig.1. As it is already brought out in previous sections that selection of an appropriate structuring element is very important step which has large bearing on the results that are produced. 3. The New Approach for Noise Removal cum Edge Detection In the proposed method, closing then followed by opening is performed using an appropriate Structuring Element (SE) on the image to be processed. Again closing operation is performed on the resultant image. This removes the noise from the image and hence is used to pre-process the image. The choosing of structuring element is a key factor in morphological image Fig.1

210 189 Fig.2 Fig.6 Fig.7 The results obtained using proposed Method-3 are shown below in Fig.8 and Fig.9. Fig.3 The results obtained using proposed Method-1 are shown below in Fig.4 and Fig.5. Fig.8 Fig.9 Fig.4 Fig.5 It could be seen from the above results that noise is completely removed from the images and edges are extracted. In Method-1, edges are thinner and some interior edges are not detected effectively in gray scale image. In Method-2, the edges are a little thicker and more number of edges could be seen. Method-3 outperforms previous two methods, edges are more continuous and almost all edges could be extracted. The results obtained using proposed Method-2 are shown below in Fig.6 and Fig Acknowledgments The authors would like to express their gratitude to Sri K.V.V.Satya Narayana Raju, Chairman and K.Sashi Kiran Varma, Managing Director, Chaitanya group of

211 190 Institutions for providing necessary infrastructure. We would like to thank Dr.V.Vijaya Kumar, Dean, G.I.E.T, Rajahmundry for his continuous support in carrying out this work and helpful discussions that led to improvise the presentation quality of this paper. We also thank the anonymous reviewers for their valuable comments. References [1] Rafael C. Gonzalez and Richard E. Woods, Digital Image Processing, Pearson Prentice Hall, [2] Milan Sonka, Vaclab Hlavac and Roger Boyle, Image Processing, Analysis and machine vision, PWS Publishers,2007. [3] Zhong Qiyuan, Huang Xianxiang, Tan Lilong and Zhau Bing, A method of deleting noise in a binary image based on the mathematical morphology, ICEMI [4] Deng Ze-Feng, Yin Zhou-Ping and Xiong You-Lun, High probability impulse noise removing algorithm based on mathematical morphology, IEEE Signal processing letters, Vol 14, No. 1, Jan [5] Yuqian Zhao, Weihua Guy and Zhencheng Chen, Edge detection based on multi-structure elements morphology Proceedings of the sixth world congress on intelligent control and automation, June 21-23, 2006, Dalian, China. [6] Yanlei Xu, Jiyin Zhao and Yubin Jiao, Noisy image edge detection based on multi-scale and multi-structuring element order morphology transformation, IEEE computer society. Principal for Vidya Vikas College of Engineering, Hyderabad for two years. Then he worked as Principal for Chaitanya Institute of Science and Technology, JNTU, Kakinada, India for one year. Presently he is working as Principal in Chaintanya Institute of Engineering and Technology, JNTU, Rajahmundry, India. He is an advisory member for many Engineering colleges. He has published 25 research articles. Presently he is guiding 10 research scholars. He is a life member of ISTE and CSI. J Sree Devi received her B.Tech (CSE) degree from JNTU, Hyderabad in She is pursuing her M.Tech from JNTU (H). She is working in M.G.I.T, Hyderabad as programmer for the last 10 years. Her research interests include Digital Watermarking, Image Processing and Pattern Recognition. AUTHORS PROFILE M. Rama Bai received B.E (CSE) degree from Bharathiar University, Coimbatore (T.N), India in She worked as lecturer in Amrita Institute of Technology and Science, Coimbatore for three years and Sri Venkateshwara Institute of Science and Technology, Machilipatnam for two years. She joined as Assistant Professor in the Dept of Computer Science & Engineering, Mahatma Gandhi Institute of Technology (MGIT) under JNT University, Hyderabad, India in She received her M.Tech (CSE) from College of Engineering, Osmania University, Hyderabad. At present she is working as Associate Professor in CSE Dept at MGIT, Hyderabad. She is pursuing her Ph.D from JNT University, Kakinada, India in Computer Science. Her research interests include Image processing, Pattern Recognition and Database Management Systems. She is a life member of ISTE. Dr V Venkata Krishna received the B.Tech. (ECE) degree from Sri Venkateswara University. He completed his M. Tech. (Computer Science) from JNT University. He received his Ph.D in Computer Science from JNT University in He worked as Professor and Head for ten years in Mahatma Gandhi Institute of Technology, Hyderabad. After that he worked as

212 191 Modeling ODP Policies by using event-b Belhaj Hafid, Bouhdadi Mohamed and Elhajji Said Department of Mathematics & Computer Science, University Mohammed V, Faculty of science BP 1014 RP, 4. Av Ibn Batouta Agdal, Rabat, Morocco Abstract The Reference Model for Open Distributed Processing (RM- ODP) defines a framework for the development of Open Distributed Processing (ODP) systems in terms of five viewpoints: information, enterprise, computational, technology and engineering. Each viewpoint language defines concepts and rules for specifying ODP systems from the corresponding viewpoint. The enterprise viewpoint focuses on the roles and policies on the enterprise that the system is meant to support. The use of formal methods in the design process of ODP systems is explicitly required. Formal notations provide precise and unambiguous system specifications. An important point to take into account is the incorporation of the many proofs which have to be performed in order to be sure that the final system will be indeed correct by construction. The Event-B method is being defined as a formal notation. In this paper, we explore the benefits provided by using the proof construction approach to specify open distributed System in the enterprise viewpoint focusing on the specification of actions and the behavioral policies conditioning them. Keywords: RM-ODP, Enterprise Language, Policies, event B, RODIN platform. 1. Introduction The RM-ODP[1,2,3,4] framework is increasingly being used for modelling complex open distributed systems, such as those in the domains of telecommunications, finance, education and defence. While some of the ODP viewpoint languages, in particular computational and engineering are developed in sufficient details to describe programming and infrastructure artifacts of any distributed system, this is not true of the enterprise language. On the other hand, the availability of maturing distributed infrastructure platforms such as CORBA, DCOM, DCE and Java-RMI increasingly encourages the use of distributed objects for business applications. As a result, the IT community is shifting its interest from platform issues towards enterprise specifications. There is an increasing demand from industry to use enterprise specifications to improve the accuracy of the design of distributed systems, in particular those that cross various administrative and organisational boundaries [23]. The existing ODP enterprise language, consisting of a limited number of concepts and structuring rules, needs further extensions and refinements in order to be better suited for enterprise modelling of practical open distributed systems. For example, there is a need for rigorous specification of policies governing the behaviour of complex systems and automated sub-systems. These policies need to be made explicit because their monitoring and enforcement will require actions by the system implemented, and the correctness of these actions can only be guaranteed if there is a well defined framework for the description of concepts such as ownership, right, objective, authority, delegation and policy. The languages Z, SDL, LOTOS, and Estelle are used in RM-ODP architectural semantics part [4] for the specification of ODP concepts. However, no formal method is likely to be suitable for specifying every aspect of an ODP system. Elsewhere, we used the meta-modeling approach [5] [6] to define syntax of a sub-language for the ODP QoS-aware enterprise viewpoint specifications. We defined a metamodel semantics for structural constraints on ODP enterprise language [7] using UML and OCL. We also used the same meta-modeling and denotation approaches for behavioral concepts in the foundations part and in the enterprise language [8,9]. Furthermore, for modeling ODP systems correctly by construction, the current testing techniques [10,11] are not widely accepted. In a previous work [12,14], we specify the trading function and the protocol of negotiating QoS requirements between enterprise objects in event B. For modeling business requirements and systems we will use the concepts provided by the RM-ODP enterprise viewpoint [15]. The enterprise viewpoint focuses on the purpose, scope and policies for the system and its environment. It describes the business requirements and how to meet them, but without having to worry about other system considerations, such as particular details of its implementation, or the technology used to implement the system. Specifically, this paper focuses on a subset of the enterprise concepts, namely on the notions of action and policy (permissions, prohibitions and obligations), and try

213 192 to provide a more precise framework for reasoning about these fundamental enterprise concepts. It aims to allow the unambiguous specification of enterprise requirements. we use the event-b formalism as our formal framework for developing policies in enterprise viewpoint ODP language. Event B [16] is a method with tool support for applying systems in the B method. Hence we can benefit from the useful formalism for reasoning about distributed systems given by refinement techniques and from the tool support in B. The Rodin Platform for Event-B provides effective support for refinement and mathematical proof. [17] The structure of this document is as follows. First, Sections 2 and 3 serve as brief introductions to the RM- ODP and event B, respectively. Then, Section 4 describes our proposal to write enterprise policies specifications in event B. Finally, Section 5 draws some conclusions and describes some future research activities. 2. RM-ODP enterprise language 2.1 RM-ODP Distributed systems are inherently complex, and their complete specifications are so extensive that fully comprehending all their aspects is a difficult task. To deal with this complexity, system specifications are usually decomposed through a process of separation of concerns to produce a set of complementary specifications, each one dealing with a specific aspect of the system. Specification decomposition is a well-known concept that can be found in many architectures for distributed systems. In particular, the Reference Model for Open Distributed Processing (RM-ODP) [1-4] provides a framework within which support of distribution, networking and portability can be integrated. It consists of four parts. The foundations part [2] contains the definition of the concepts and analytical framework for normalized description of arbitrary distributed processing systems. These concepts are grouped in several categories which include structural and behavioral concepts. The architecture part [3] contains the specifications of the required characteristics that qualify distributed processing as open. It defines a framework comprising five viewpoints, five viewpoint languages, ODP functions and ODP transparencies. The five viewpoints are enterprise, information, computational, engineering and technology. Each viewpoint language defines concepts and rules for specifying ODP systems from the corresponding viewpoint. However, RM-ODP is a meta-norm [5] in the sense that it defines a standard for the definition of other ODP standards. The ODP standards include modelling languages, specification languages and verification[12, 13]. 2.2 The Enterprise Viewpoint RM-ODP [1-4] provides five generic and complementary viewpoints on the system and its environment: enterprise, information, computational, engineering and technology viewpoints. They enable different abstraction viewpoints, allowing participants to observe a system from different suitable perspectives [18]. The enterprise viewpoint focuses on the purpose, scope and policies for the system [15] and its environment. It describes the business requirements and how to meet them, but without having to worry about other system considerations, such as particular details of its implementation, or the technology used to implement the system. Bellow, we summarize the basic enterprise concepts. Community is the key enterprise concept. It is defined as a configuration of enterprise objects formed to meet an objective. The objective is expressed as a contract that specifies how the objective can be meet[19]. A contract is a generic concept that specifies an agreement governing part of the collective behavior of a set of objects. A contract specifies obligations, permissions and prohibitions for objects involved. The scope of the system is defined in terms of its intended behavior, and this is expressed in terms of roles, processes, policies, and their relationships. Roles identify abstractions of the community behavior, and are fulfilled by enterprise objects in the community. Processes describe the community behavior by means of (partially ordered) sets of actions, which are related to achieving some particular sub-objective within the community. Finally, policies are the rules that constrain the behavior and membership of communities in order to achieve their objectives. A policy can be expressed as an obligation, a permission, or a prohibition. Obligation: A prescription that a particular behaviour is required. An obligation is fulfilled by the occurrence of the prescribed behaviour (RM-ODP, part 2, clause ). Permission: A prescription that a particular behaviour is allowed to occur. A permission is equivalent to there being no obligation for the behaviour not to occur (RM- ODP, part 2, clause ). Prohibition: A prescription that a particular behaviour must not occur. A prohibition is equivalent to there being an obligation for the behaviour not to occur (RM-ODP, part 2, clause ). In general, ODP systems are modeled in terms of objects. An object is a model of an entity; it contains information and offers services. A system is therefore composed of interacting objects. In the case of the enterprise viewpoint we talk about enterprise objects, which model the entities defined in an enterprise specification [20].

214 193 An enterprise object is an object that filles one or more roles in a community. It can also participate in more than one community at one time. An enterprise object may be a role, an activity or a policy of the system. [19] event in the abstract model is composed of a guard and an action. A typical abstract machine may be outlined as below. MACHINE M SETS S1,S2,S3... CONSTANTS C PROPERTIES P VARIABLES v1,v2,v3... INVARIANTS I INITIALISATION init EVENTS E1 = WHEN G1 THEN S1 END; END. 4. Specifying ODP policies by using event B 4.1 Abstract and concrete levels on enterprise concepts Fig. 1 Enterprise concepts [19]. 3. Event B modeling approach The Event-B [21] [22] is formal techniques consist of describing rigorously the problem, introduce solutions or details in the refinement steps to obtain more concrete specifications and verifying that proposed solutions are correct. The system is modeled in terms of an abstract state space using variables with set theoretic types and the events that modify state variables. Event-B, a variant of B, was designed for developing distributed systems. In Event-B, the events consist of guarded actions occurring spontaneously rather than being invoked. The invariants state properties that must be satisfied by the variables and maintained by the activation of the events. The mathematical foundations for development of event based system in B is discussed in [13]. An abstract machine consists of sets, constants and variables clause modelled as set theoretic constructs. The invariants and properties are defined as first order predicates. The event system is defined by its state and contain number strained by the conditions defined in the properties and invariant clause known as invariant properties of the system. Each The interaction of people with IT systems generate various restriction needs to guarantee that each system user benefits of its advantages without trespassing on another user s rights. These needs vary according to the activity field required. It could be regarding: Confidentiality (Non disclosure of sensitive information to non authorised persons), Integrity (Non alteration of sensitive information), Availability (Supply of information to users according to their rights of access these information), Auditability (The ability to trace and determine the actions carried out in the system). Such requirements usually result in expressing policies, defining for each user his permissions, prohibitions and obligations. Users (or objects type) are active entities operating on enterprise objects (passive entities) of the system. Summing up, an enterprise specification is composed of specifications of the elements previously mentioned, i.e. the system s communities (sets of enterprise objects), roles (identifiers of behavior), processes (sets of actions leading to an objective), policies (rules that govern the behavior and membership of communities to achieve an objective), and their relationships [15]. A contract specifies obligations, permissions and prohibitions for objects comprising in a communities. Just as for the objects, the actions are also gathered in processes, this implies that there are two levels of abstraction in ODP enterprise viewpoint: Abstract level: roles, processes and enterprise viewpoint of the system on which various permissions, prohibitions and obligations are expressed. Concrete level: object type (client, server, policy maker, policy administrator), actions (create, delete) and enterprise objects of the system.

215 194 Object type, actions and enterprise objects are respectively assigned to roles, processes and enterprise viewpoint by relations defined over these entities(see figure 2). We detail relations in the next sub-section. Play, Use and belong. Assignment of Objects type to roles: Objects type are assigned to one or more roles in order to define their privileges. Objects type play their roles in communities, which implies that these objects are assigned to roles through a ternary relation including the community: play(com, Ot, r): means that the Object type Ot plays the role r in the community com. Assignment of actions to processes: As for roles and Objects type, processes are an abstraction of various actions authorized in the system. The relation binding actions to processes is also a ternary relation including the community: belong(com, a, p): means that the action a is considered as a process p in the community com. Assignment of enterprise objects to enterprise viewpoint: The relation binding the enterprise objects to the enterprise viewpoint to which they belong is also a ternary relation including the community: use(com, o, v): means that the community com uses the object o in the enterprise viewpoint v. Permission Obligation (com, role, process, enterprise viewpoint,) Play Belong Use Prohibition Abstract level Concrete level (object type, action, enterprise object ) Fig. 2 Abstract and Concrete level of ODP enterprise viewpoint s concepts. Modeling a Policy. When objects type, actions, and enterprise objects are respectively assigned to roles, processes and enterprise viewpoint, it is now possible to describe the policy. It consists of defining different permissions, prohibitions and obligations: permission(com, r, p, v): means that the community com grants to the role r the permission to carry out the process p on the enterprise viewpoint v. prohibition(com, r, p, v): means that the community com prohibits the role r to carry out the process p on the enterprise viewpoint v. obligation(com, r, p, v): means that the community com require the role r to carry out the process p on the enterprise viewpoint v. Hierarchy In situations when two or more groups of objects, under control of different autorities, engage in cooperation to meet a mutual objective, they form a specifal kind of community called a federation. The hierarchies allow the inheritance of the privileges (permissions prohibitions or obligations), if for example r2 is a sub-role of r1, for a community com, a process p and an enterprise viewpoint v: When permission(com, r1, p, v) holds then permission(com, r2, p, v) holds. When prohibition(com, r1, p, v) holds then prohibition(com, r2, p, v) holds. When obligation(com, r1, p, v) holds then obligation(com, r2, p, v) holds. In the same way for the communities, if com2 is a subcommunity of com1 then, for a role r a process p and an enterprise viewpoint v: When permission(com1, r, p, v) holds then permission(com2, r, p, v) holds. When prohibition(com1, r, p, v) holds then prohibition(com2, r, p, v) holds. When obligation(com1, r, p, v) holds then obligation(com2, r, p, v) holds Event B Models for ODP policies The expression of the policy in event B includes several successive stages. A first B model is built and then other successive refinements are made as shown by figure 3. The first refinement validates the link between the abstract level (role,...) and the concrete level (object type,...). The approach is based on refinement and each model or refinement model is enriched either by constraints required by the system specification. Each constraint is attached to an invariant. The invariant becomes stronger through the refinement steps Abstract Model with policies As presented in the paragraph 4.1, the enterprise specification has two levels of abstraction (see figure 2). The first step consists of an event B model modeling the abstract part of the policy, i.e. initially, only concepts of community, role, enterprise viewpoint, process are considered. In the first model, permissions, obligations and prohibitions should be described. The clause SETS in the event B model contains basic sets such as community, roles, processes, enterprise viewpoint: COMS, ROLES, PROCESSES, ENT VP.

216 195 ODP Policy B Abstract model with permissions, obligations and prohibitions refinement B Concrete model with permissions, obligations and prohibitions refinement Continue to develop the system with B Fig. 3. Steps of conception of event- B based model of ODP policies The clauses CONSTANTS and PROPERTIES contain the constants like permission, obligation and prohibition that will contain privileges of the ODP system description. Two new constants sub_role and sub_com are introduced to take into account respectively the role and community hierarchy. It is enough to specify which roles and which communities are concerned with inheritances, and the permissions, obligations and prohibitions corresponding to inheritances are deductively generated. SETS COMS; ROLES; ROCESSES;ENT VP; Checking Consistency Specifying policies with event B CONSTANTS permission, prohibition, obligation, sub_com, sub_role PROPERTIES permission COMS ROLES PROCESSES ENT VP prohibition COMS ROLES PROCESSES ENT VP obligation COMS ROLES PROCESSES ENT VP sub_org ORGS ORGS sub_role ROLES ROLES / * Organization hierarchies * / (com1, com2, r, p, v). ((com1 COMS com2 COMS r ROLES p PROCESSES v ENT VP (com1 com2) sub_com (com2 r p v) permission) (com1 r p v) permission) / * Role hierarchies * / (com, r1, r2, p, v). ((r1 ROLES r2 ROLES com COMS p PROCESSES v ENT VP (r1 r2) sub_role (comr2pv)permission) (com r1 p v) permission) / * Same properties for prohibitions and obligation * / For a given particular case, it is enough to initialize sets in the clause SETS by entities, communities, roles, enterprise viewpoint, processes. Properties of constants, like permission, prohibition, obligation, sub_role and sub_com, should also be set in the clause PROPERTIES. Consequently, permissions, prohibitions and obligations cannot be modified, since they are defined as constants. Introducing State Variables. An event B model expresses properties over state and state variables. Variables are used to model the status of the system with respect to permissions, prohibitions and obligations: The clause VARIABLES contains the state variable hist_abst that contains the history of system processes and satisfy the following properties added to the invariant: INVARIANT hist_abst COMS ROLES PROCESSES ENT VP hist_abst permission The initial values of the variable is set as follows: hist_abst := As the policy is supposed to be consistent, we should be able to prove in the clause ASSERTIONS : ASSERTIONS permission prohibition= permission obligation = hist_abst prohibition = hist_abst obligation = The clause EVENTS contains the following event : The event action models when an authorization request for the access of an object type to an enterprise object of the system occurs. action any com, r, v, p where com COMS r ROLES v ENT VP p PROCESSES (org r p v) permission then hist_abst := hist_abst {(com r p v)} end The invariant should be preserved and it means that any process in the system is controlled by the policy through the variable hist_abst First Refinement: Concrete Model with Policies We defined two levels of abstraction and the current model is refined into a concrete model. The refinement introduces object types, actions and enterprise objects: sets OBJTYPES, ACTIONS and ENTOBJ contain respectively object types, actions and enterprise objects of the system under development. The clause CONSTANTS includes the following constants: play (assignment of objects types to roles), use (assignment of objects to enterprise viewpoint) and belong (assignment of actions to processes). Properties of constants are stated as follows: PROPERTIES play COMS ROLES OBJTYPES use COMS ENT VP ENTOBJ

217 196 belong COMS PROCESSES ACTIONS Concrete Variables. A new variable hist_conc models the control of the system according to the policy; it contains the history of the actions performed by an object type on a given enterprise object. The context in which the action occurred is also stored in this variable. The relation between hist_conc and the variable hist_abst of the abstract model is expressed in the gluing invariant; the first part of the invariant states properties satisfied by variables with respect to permissions. INVARIANT (ot, a, o).( (ot OBJTYPES a ACTIONS o ENTOBJ (ot a o ) hist_conc) ((com, r, p, v).(com COMS r ROLES p PROCESSES v ENTVP (r ot) play (v o) use (p a) belong (com r p v) hist_abst))) The invariant states that each action performed by the system satisfies the policy. For the prohibitions, when a subject s wants to carry out an action a on an object o in an organization org, it is necessary to check that no prohibition exists for that action. The second part of the invariant states properties satisfied by variables with respect to prohibitions and obligations: INVARIANT (ot, a, o).( (ot OBJTYPES a ACTIONS o ENTOBJ (ot a o) hist_conc) ((com, r, p, v).(com COMS r ROLES p PROCESSES v ENTVP (r ot) play (v o) use (p a) belong) (com r p v) prohibition) (com r p v) obligation)) action any ot, a, o, com, r, v, p where ot OBJTYPES a ACTIONS o ENTOBJ com COMS r ROLES p PROCESSES v ENTVP (r ot) play (v o) use (p a) belong / permission / (com r p v) permission / prohibition and obligation / ((comi, ri, pi, vi).((comi ORGS ri ROLES pi PROCESSES vi ENTVP (ri ot) play (vi o) use (pi a) belong) ((comi ri pi vi) prohibition) (comi ri pi vi) obligation)) Then hist_conc := hist_conc {(ot a o)} end The Events. The abstract model should consider the permissions, the prohibitions and the obligations for an object type ot that asks to perform an action a on an enterprise object o. 5. Conclusions The use of formal methods in the design process of ODP systems is explicitly required. An important point to take into account is the incorporation of the many proofs which have to be performed in order to be sure that the final system will be indeed «correct by construction». In this article We presented our approach for developing distributed system in Event B. we used event B for modeling policies in ODP enterprise viewpoint. The work was carried out on the Rodin platform. In order to verify our models, the abstract and refinement model of ODP policies are developed by using Event-B, Each model is analyzed and proved to be correct. Our experience strengthens our believe that abstraction and refinement are valuable technique for modeling complex distributed system. As for future work, we are going to generalize our approach to other concepts in ODP systems. This will be our basis for further investigation of using event-b in the design process of ODP systems. Moreover, case studies should be developed using these models. References [1] ISO/IEC, Basic Reference Model of Open Distributed Processing-Part1: Overview and Guide to Use, ISO/IEC CD , 1994 [2] ISO/IEC, RM-ODP-Part2: Descriptive Model, ISO/IEC DIS , [3] ISO/IEC, RM-ODP-Part3: Prescriptive Model, ISO/IEC DIS , [4] ISO/IEC, RM-ODP-Part4: Architectural Semantics, ISO/IEC DIS , July [5] M. Bouhdadi and al., A UML-Based Meta-language for the QoS-aware Enterprise Specification of Open Distributed Systems IFIP Series, Vol 85, Springer, (2002) [6] Mohamed Bouhdadi and al. A Semantics of Behavioural Concepts for Open Virtual Enterprises. Series: Lecture Notes in Electrical Engineering,, Vol. 27.Springer, p [7] Belhaj H and al. Event B for ODP Enterprise Behavioral Concepts Specification, Proceedings of the World Congress on Engineering 2009 Vol I, WCE '09, July 1-3, 2009, London, U.K., Lecture Notes in Engineering and Computer Science, pp , Newswood Limited, 2009 [8] Mohamed Bouhdadi and al., Using BPEL for Behavioural Concepts in ODP Enterprise Language, Virtual Enterprises and Collaborative Networks, IFIP, Vol. 283, pp , Springer, 2008 [9] Mohamed Bouhdadi and al., Meta-modelling Syntax and Semantics of Structural Concepts for Open Networked Enterprises, Lecture Notes in Computer Science, Vol. 4707, pp , Springer, [10] Myers, G. The art of Software Testing, John Wiley &Sons, New York, 1979 [11] Binder, R. Testing Object Oriented Systems. Models. Patterns, and Tools, Addison-Wesley, 1999

218 197 [12] Belhaj Hafid and al.: Using Event B to specify QoS in ODP Enterprise language. PRO-VE'10 11th IFIP Working Conference on VIRTUAL ENTERPRISES, Saint-Etienne, France, October [13] J.-R. Abrial. The B-Book: Assigning programs to meanings. Cambridge University Press, [14] Belhaj Hafid, Bouhdadi Mohamed, El hajji Said : Verifying ODP trader function by using Event B. IJCSI International Journal of Computer Science Issues, Vol. 7, Issue 4, No 9, July [15]ISO/IEC. RM-ODP Enterprise Language. Draft International Standard ISO/IEC 15414, ITU-T X.911, ISO, [16] [17] RODIN. Development Environment for Complex Systems (Rodin) [18] P. Linington. RM-ODP: The architecture. In K. Milosevic and L. Armstrong, editors, Open Distributed Processing II, pages Chapman & Hall, Feb [19] Mohamed Bouhdadi and al. A UML/OCL Meta-model Syntax for Structural Constraints in ODP Enterprise Language Journal WSEAS Transactions on Computers, Vol 6, Issue 1, WSEAS Press, pp:31-36, [20] Francisco Duran, Javier Herrador, and Antonio Vallecillo. Using UML and Maude for Writing and Reasoning about ODP Policies. Proceedings of the 4th International Workshop on Policies for Distributed Systems and Networks (POLICY 03), Lake Como (Italy). pp , IEEE Computer Society Press, June [21] Joochim, T., Snook, C., Poppleton, M. and Gravell, A. (2010) TIMING DIAGRAMS REQUIREMENTS MODELING USING EVENT-B FORMAL METHODS. In: IASTED International Conference on Software Engineering (SE2010), February 16 18, 2010, Innsbruck, Austria. [22] C.Snook & M.Butler, UML-B and Event-B: an integration of languages and tools. Proc. IASTED International Conf. on Software Engineering (SE2008), Innsbruck, Austria, [23] P.F. Linington, Z Milosevic, and K. Raymond, Policies in communities: Extending the enterprise viewpoint. In Proc. 2nd International Workshop on Enterprise Distributed Object Computing (EDOC'98), San Diego, USA, page 11, November 1998.

219 198 A New Security Paradigm of File Sharing Seifedine Kadry Faculty of General Education American University of the Middle East Abstract Windows Right Management Services protects RMS-enabled files or applications from unauthorized users. However, the offered security on the whole file prevents other trusted recipients with minor privileges to access it. The sender is obliged to send each time to these recipients another file that resembles to the former but does not contain sensitive information which is considered as time wasting especially when the number of recipients is increased. This paper designs and implements a new security layer that extends the WRMS security provided on a certain file, in a way that the file still keeps its security towards unintended or unauthorized recipients, but can be sent only once to trusted recipients having different privileges in such a way that each recipient will only see the data that grants access to. Keywords: WRMS, File Sharing, Security, XML, XrML,.NET. 1. Introduction Windows Right Management Services proves its ability in protecting RMS-enabled information from unauthorized use no matter where this information goes inside or outside the organization [1]. This solution comes subsequent to many perimeter-based methods such as Firewalls, Access Control Lists, encryption, and authentication technologies. Although these methods could protect the information while in transit, but couldn t provide any layer of security when the information is exposed to the recipient. WRMS was the first technology that could introduce this layer of security whenever the file is received because the permission to the file is stored with the file itself. It can prevent the recipient to copy, modify, print, forward the RMS-based file via to non-intended or perhaps malicious recipients, or to access it after a certain period of time. However, WRMS also prevents other trusted recipients with minor privileges to access a certain file since the offered security is implemented on the whole document [2]. Thus, the sender is obliged each time to send to these recipients another file similar to the former but does not contain confidential data. This is considered as time consuming especially when we are talking about a big organization that has numerous branches perhaps in different countries and each branch consists of many departments. This limitation of WRMS motivated us to think about a solution to the addressed problem. The objective of this paper is to extend the functionality of WRMS by adding a new security layer to Microsoft Excel worksheets, in such a way that the file still keeps its WRMS features and its security towards unintended recipients, but can be sent only once to trusted recipients having different privileges. To allow to this file to be sent to different recipients, the provided solution will apply a security on a portion of the document, this portion contains the sensitive data. As a final result, each recipient will only see the data that grants access to. This paper is divided into the following sections: In section two, we will discuss why we need this solution. And we will explain the importance of WRMS and how it works. The proposed solution that we adopt in this paper and how it is implemented, are clearly detailed in section three. Section four will discuss the conclusion, and the advantages of this solution in addition to future works in this domain. 2. Problematic For a better understand the goal of this paper, we must look at the described real life scenario: At the end of the business year, the Financial Director of GlobalCom Company prepares the annual bonuses file. The company has three departments: IT, Logistics and Quality Assurance. Each department has a manager and a group of employees that reports to this manager. Every department s manager receives from the financial director the bonuses file and approves it. The confidentiality plays an important role here since it is part of the company policy. For instance, the IT manager is not allowed to view the Logistics and Quality bonuses. The same scenario is repeated for the Logistics and Quality department, taking into consideration that all departments are sharing the same file. How to achieve this? The financial director has a program which will allow him to encrypt any data from the Excel

220 199 file. In our case, the bonus column value for all employees in all departments will be the encrypted data. In other words, the role of the sender is to encrypt whatever data from the excel sheet and sends it to each department s manager, and to apply WRMS to this file to assure that the file will not be edited or forwarded. The receiver s role is to view only its authorized data. In absence of our solution: the final director will create three files to be sent for each department s manager and applies WRMS on each of these files without getting the benefit to send the common data. Each of the received RMS-enabled file will contain the bonus value to each department. In the presence of our solution: The financial director can get the benefit of the file-reusability and will send a single shared RMS-enabled file to each department s manager. This file contains encrypted bonuses column for all employees in all departments. Upon receiving the file by each department s manager, only the relevant data of each department will be decrypted, and the rest of the data that is relevant to other department s manager will remain encrypted because this solution allows the decryption of selective information based on the privileges of each department s manager. After presenting both cases before and after applying the solution, it is logical to adopt this study that addresses this problem. It is obvious that this solution is time and resource effective (file reusability) and it deals with confidential data of the company in a secure way up to the latest technology techniques. 2.1 What is WRMS? Windows Right Management Services is information protection technology that works with RMS-enabled applications to protect digital information from unauthorized use both online and offline, inside and outside of the firewall. RMS increases the security strategy of an organization by providing protection of the information through usage rights and conditions, which remain with the information, no matter where it goes. In other words, these permissions are assigned to an authorized recipient after the information is accessed [3]. 2.2 WRMS Workflow In a reference to [3], Windows RMS which includes both server and client components provides the following capabilities: Creating rights-protected files and containers: Users who are trusted entities in an RMS system could apply usage rights and conditions to digital information using RMS-enabled applications or browsers. They can easily create and manage protected files using applications that people use every day -for example: computer-aided design (CAD) applications or Microsoft Office 2007 Editions- that incorporate Windows RMS technology features. Using common task management procedures within a familiar onscreen environment, organizations could assign usage rights and conditions to digital information such as an message or document. In addition, RMS-enabled applications provide users with the option of applying authorized rights policy templates such as Company Confidential. Licensing and distributing rights-protected information: The XrML-based [4] certificates issued by an RMS system (right account certificate) identify trusted entities that can publish or view rightsprotected information. Users who are trusted entities in an RMS system can assign usage rights and conditions to information they want to protect through an RMS-enabled application. These usage policies specify who can use the information and what they can do with it. The RMS system validates transparently the trusted entities and issues the publishing licenses that contain the specified usage rights and conditions for the information. The information is encrypted using the electronic keys from the RMS-enabled application (for example: Microsoft Word) and the XrML-based certificates of the trusted entities. After the information is encrypted or locked, only the trusted entities specified in the publishing licenses can unlock and use that information. Users could then distribute the rights-protected information to other users in their organization via , internal servers, or external sites to enable trusted external partners to access the information. Acquiring licenses to decrypt rights-protected information and enforcing usage policies: Trusted entities recipients who are named by information author, can open or view rights - protected information by using trusted computers having WRMS client software

221 200 installed along with RMS-enabled applications or browsers. In a process that is transparent to the recipient, the RMS server, which has the public key that was used to encrypt the information, validates the recipient and then issues a user license that contains the usage rights and conditions that were specified in the publishing license. The information is decrypted using the electronic keys from the end-user license and the XrML-based certificates of the trusted entities. The usage rights and conditions are then enforced by the RMS-enabled application. The usage rights are persistent and enforced everywhere the information goes. Referring to [3], for a better understanding of the RMS workflow within an organization, see Figure 1: Fig1: RMS Workflow 1. Authors receive a client licensor certificate from the RMS server the first time they rights-protect information. (This is a one-time step that enables offline publishing of rights-protected information in the future.) 2. An author creates a file and defines a set of rights and rules. The RMS-enabled application in conjunction with the Windows Rights Management client software creates a publishing license and encrypts the file. 3. The author can distribute the file in any preferred manner. 4. The recipient clicks the file to open it. The RMSenabled application calls to the RMS server, which validates the user and issues a user license. The application renders the file and enforces the rights defined in the use license. 3. The Proposed Solution 3.1 Design Since WRMS works in Microsoft-based environment, it was logical to use VB.NET [5] in our solution as an object oriented programming language in order to get some benefits of.net Framework class libraries. The proposed solution should be tested first in VB.NET Windows Form Application and then it should be applied in Excel application where there is a need to be created in a separate module. This security layer provides us the opportunity to encrypt a simple node or an entire column through the.net and XML technology. The implementation of this solution in VB.NET in the Encryption process includes: Converting Excel to XML file, Encrypt / Decrypt XML Nodes, and converting the encrypted XML to Excel file (Figure 2). The conversion from an Excel to XML file helps us to select the specific nodes to be encrypted in later step. XML was used because it is the most popular technology for structuring data; therefore XML-based encryption is the natural way to handle complex requirements for security in data interchange applications [8]. In the decryption process, the result is revealed: The description of the Microsoft Windows logged-on user is compared with the privileges of each user in the Excel input file. Thus, each recipient can see the information that grants access; the other part of the data will remain encrypted. The decrypted XML file is also converted to decrypted Excel file. Finally, after testing the proposed solution in VB.NET Windows Form Application, we can apply it in Excel file supporting macros and we must get the same result.

222 Implementation This code in the proposed solution is divided into two main parts: A. Windows Form Application: Excel Security in VB.NET Convert Excel File to XML: Fig2: Workflow of the proposed solution 3. Selecting the data from the Excel file using DataAdapter object. 4. TableMapping. 5. Fills this DataTable with the imported data. 6. Set the property Table with the first DataTable contained in the DataSet DtSet. 7. Exports this Dataset to an XML file. 8. Reads the DataColumns and adds them the ArrayList collist. 9. Closes the connection to the data source. ADO.NET: stands for (ActiveX Data Object), is special set of.net framework allows the user to work with different type of databases such as Access, SQL server and Oracle. It provides two ways to work with data in a database: connected mode and disconnected mode [5]. In order to convert Excel file to XML file, we will use the ADO.NET in disconnected mode. The ConvertExcelToXML()method will perform the following Actions: 1. Creates a connection to the Excel file using OLEDBConnection object. 2. Opens a connection with the property settings specified. To create a connection to the Excel file: Referring to [6, 7], we have three ways to manipulate an Excel file. It can be done either by using Microsoft Office Component, Microsoft Jet Engine, and Access Database Engine. As per Microsoft recommendation, it is not advisable to use Office components on the server. Since Microsoft Jet Engine is only used to open a connection to Excel 2003 worksheets, but doesn t support connectivity to Excel 2007 worksheets. So, the connection will be done using Microsoft.ACE.OLEDB.12.0 engine. Before using Access Database Engine we must download the 2007 Office system driver from:

223 ID=7554F536-8C B72- EF94E038C891&displaylang=en To connect to an Excel file, we can use an OLEDB object that treats Excel as a database. The required information can be easily fetched by using SQL queries. 202 To select an ordered list xmllist of XML nodes, we should create an object of XmlNodeList() and use the function SelectNodes()which takes an argument as X-Path expression. Referring to [9], X-path is a syntax for defining parts of an XML document. Encrypt Node(s) in XML file: Since we need to encrypt only a portion of the file, XML Encryption is the best choice if the application requires a combination of secure and insecure communication (which means that some of the data will be securely exchanged and the rest will be exchanged as is)[8]. The SelectNodesToEncrypt()method will perform the following actions: 1. Gets the EmployeeID of the nodes to be encrypted. 2. Loops through the selected items and creates an EncryptedXML object exml. 3. Selects the XML element(s) needed to be encrypted. 4. Encrypts the element(s) using the key generated sharedkey by the encryption algorithm object. 5. Creates an encrypted data object ed and specifies its properties. 6. Creates a cipherdata element and sets its value to the encrypted XML element EncryptedElem. 7. Replaces the plaintext XML element Elem with the encrypted data object ed. 8. Saves the encrypted data to a file "EncryptedData.xml" Convert Encrypted XML file to Excel: The ConvertXMLtoExcel()will perform the following actions: 1. Creates a new Excel Application, workbook exbook containing one Datasheet exsheet. 2. Creates a new dynamic DataSet which has a DataTable that itself contains DataRows and DataColumns. 3. Loads the encrypted XML file specified by the user. 4. Creates DataColumns, sets their Data types and adds them to the DataTable. 5. Creates XMLNodeList object emplist that contains all TableMapping nodes specified by the user. 6. The DataRows of the XMLElements in emplist are filled with data available in EncryptedData.xml file 7. Sets the value of the property DataTableENC to the first table of the dataset ds. 8. Fills the Datasheet s cells exsheet with the DataTable s headers and rows. 9. Saves the new workbook Excel file and closes it, and quits the Excel Application. Encrypt Entire Column in XML file: The SelectColumnToEncrypt()method will perfom the following actions: This method is similar to the SelectNodesToEncrypt() method but instead, it takes the entire column to be encrypted. Therefore, an XMLNodeList object will be created to hold all the elements inside the plaintext XML file: Decrypt the XML file: The DecryptXML()method will perform the following actions: 1. Loads the encrypted XML document specified by the user. 2. Creates an XMLNodeList reglist holding all encrypted nodes in the encrypted XML document.

224 Loops through this list and searches inside all <Employee> tags if they contain the current description of the logged on user. 4. When this description is found, it retrieves the encrypted XML Element(s), Creates an encrypted data object ed2 and loads the encrypted element c.firstchild into the encrypted data object. 5. Creates an encrypted XML object exml2. 6. Decrypts the encrypted element ed2 using the shared key 7. Replaces the encrypted element c.firstchild with the plain-text XML element decryptedelem. 8. Saves the decrypted data to an external file (optional). Post XML file in Web Browser: The PostXMLFile() method will perform the following actions: It declares an instance of URI uri that takes the file name as an argument and sets the WebBrowser s URL s value to this instance. Write the shared key: The WriteSharedKey() method will perform the following actions: 1. Creates an instance sharedkey of the encryption algorithm provider 2. Converts the generated key to Base-64 String and write it to an external file specified by the user. 1. Create an instance of the encryption algorithm provider chosen: here, TripleDES will be the algorithm used to encrypt the XML data. The.NET framework provides us with multiple types of encryption algorithms we can use such as: Triple DES, AES 128, AES 192, AES 256, RSA (Rijndael algorithm) and X509CertificateEx. 2. Using the instance created sharedkey, we can get the value of the secret key for this algorithm used, and convert it from 8 bit-array to Base-64 String and write it to an external file specified by the user: Read the shared key: The ReadSharedKey() method will perform the following actions: 1. Reads the stored generated key using StreamReader class. 2. Converts the read key from base 64 digit to an equivalent 8-bit integer array. 3. Sets the secret s key value for the TDES algorithm to the convert 8-bit read key. 1. To read the stored shared key, we should create an instance reader of the StreamReader class: 2. After reading the key stream until the end, we can convert it from base 64-bit to an equivalent 8-bit integer array and saves it into the data() array 3. Set the secret s key value sharedkey to the converted 8-bit array and close the reader: Load the XML file: The LoadXmlFile()method will perform the following actions: It sets the value of the property to the new XMLDocument(), and then it loads the XML Document from the user specified location. If an error occurred during loading this file, an error message will appear. Get Current User Information: The GetCurrentUserInfo()function will perform the following actions: 1. Searches for all win32 accounts available in this local machine and saves them in ColCSItems object. 2. Gets the current logged-on user. 3. Gets the user Description for this logged-on user. 1. This function will search for all win32 user accounts [10] available in this local PC including the built-in accounts and saves them in ColCSItems object [11]. ConnectionOptions.Impersonation property sets COM impersonation level to be used for operations in this connection. In our case,

225 204 root\cimv2 is the namespace used and WIN32 is the provider used. impersonationlevel=impersonate is used when the provider is a trusted application or service. It eliminates the need for the provider to perform client identity and access checks for the requested operations [12]. 2. It also gets the current logged-on user. Since the name of the logged-on user will be returned as computername\username shape and we are only interested in the username itself, we should split it and get the second part of it and save it in a string array called name. This string array will hold both parts on the split logged-on user [10]. 3. Loops through all the win-32 accounts retrieved, when a match is found between the win-32 user and the logged-on user: we can get the description of this logged-on user and return it. Because we are only interested in the description of the logged-on user, the other cases are skipped. Now, after we have presented the SecurityLayer class, we must build it and convert it to dll file to be used as reference in both Excel Security form application and emp.xlsm Excel file Visual Basic Editor. The following items are required in order to create this form: Browse Button called btn_browse Convert encrypted XML to Excel Button called Btn_ConvertEncryptedXMLtoExcel Convert decrypted XML to Excel Button called Btn_ConvertDecryptedXMLtoExcel Decrypted XML Button called Btn_DecryptXML Instance of TabControl to contain 6 tab pages called TabControl1 Textbox to write the path of the excel file called txt_path CheckBox to select all nodes in the column called chk_all ComboBox to select which column should be encrypted called cb_colnames Instance of OpenFileDialog control called OpenFileDialog1 Three instances of DataGridView control called DataGridView1, DataGridView2, and DataGridView3 to display the plaintext Excel file, the encrypted Excel file and the decrypted files respectively. Instances of WebBrowser control to display the content of encrypted XML, Decrypted XML and plaintext XML called webbrowser1, webbrowser2 and webbrowser3 respectively. We can set the text property of the form to New Security Layer To Excel We could set the properties of both txtplaintext, txtencrypted and TxtDecryptedXML to the following: ReadOnly: True, to prevent the user to change the text s content. MultiLine: True, to display more than one line of the text. Scrollbars: Both, Both the vertical and the horizontal scrollbars are displayed when the text contains multiline. We can set an initial directory to OpenFileDialog1 control: InitialDirectory: c: so, the first directory will open when the user clicks on the browse button is C:. We set the AutoSizeColumnMode property to AllCells of the DataGridView1, DataGridView2 and DataGridView3 to determine the auto size mode for the visible columns. We must set the DropDownStyle of the cb_colnames ComboBox to DropDownList to oblige the user to select an Item from the list. To add the 6 tab members to the TabControl1, we need to click on the tabpages: (Collection) property and add 6 members TabPage1 to TabPage6, and set their texts to: XML Data, XML Plaintext, and Encrypted XML, Encrypted Excel File, Decrypted XML and Decrypted Excel File respectively.

226 205 Finally, we should add a column to the DataGridView1 to allow the user to select a node to be encrypted. This is done by clicking on the property Columns: (Collection) and add a column. The added column has SelectEmployee as a Name, DataGridViewCheckBoxColumn as a type, and Select as a Header text. After the SelectEmployee column is added to the DataGridView1, the user is able to browse the Excel file: Fig4: Displaying the Excel data in a DataGridView Fig3: Final form shape The Btn_Browse_Click()method will perform the following actions: 1. Filters the Excel file Extension choice 2. Saves the selected file in textbox control 3. Sets the values of properties in SecurityLayer class & convert the Excel file to XML 4. Sets the DataGridView s DataSource to the property datatable 5. Fills the ComboBox cb_colnames with the items in columnlist property 6. Posts the converted xml file on the xml plaintext tab and sets this DataGridView to read-only. Fig5: Displaying the exported XML file. Set the DataGridView to Read-only : The SetGridColumn() method: Since the first column of the DataGridView SelectEmployee should be editable to allow the user to select node(s) to encrypt,the SetGridColumn()sub will loop through each column of this DataGridView,except the first column, and make their values read-only in order to prevent the user from changing the DataGridView Cells:

227 206 Selecting all nodes in a column in one click: 3. User Validation 4. Checks whether chk_all" checkbox was selected, and calls either SelectColumnToEncrypt() and Posts the encrypted xml file. 5. If not, it calls SelectNodesToEncrypt() & also posts the encrypted xml file in web Browser. Fig 6: Encryption of all nodes in BASICSALARY by clicking chk_all checkbox The chk_all_checkedchanged() method: The following method assigns the value of the checkboxes in the DataGridView1 to the value of chk_all checkbox. If it was selected, then these checkboxes will be selected too; similarly if chk_all was reset, these checkboxes will be reset too. This sub also makes the value of the DataGridView1 read-only after the chk_all checkbox was checked, to prevent the user to un-check any cell in the DataGridView1. Encrypt node(s) or entire column in XML file: Fig 8: No node or an entire column was selected to be encrypted Fig 9: The Column from the ComboBox wasn t selected Fig 7: Encryption of <BASICSALARY> XML node The Btn_encrypt_click()method will perform the following actions: 1. Loads the XML document posted in XML Plaintext tab. 2. Sets the values of some properties and writes the shared key to an external file. 1. Check whether chk_all" checkbox was selected. If so, call SelectColumnToEncrypt()method and post the encrypted file in Encrypted XML tab: 2. If chk_all" checkbox was not selected, loop through the ID of the selected nodes and save them in a string separated by, to call SelectNodesToEncrypt() method. the encrypted file is also posted under Encrypted XML tab in a web Browser:

228 207 Decrypt the encrypted XML file: Convert the encrypted XML to Excel file: Fig 10: Decryption of the XML file for the Manager,B1 The Btn_DecryptXML_Click() method will perform the following actions: The values of some properties in securitylayer class should be set to call the decryptxml() method. the decrypted xml file will be posted on the Web Browser. Fig 11: Encryption of the some nodes in BASICSALARY column The Btn_ConvertEncrypted_XMLtoExcel_Click() method will perform the following actions: The values of both InputXMLFile() and OutPutExcelFile() properties will be set in order to call the ConvertXMLtoExcel() method. The result will be posted on DataGridView2 Note: Only the information that the loggedon user has access to it will be decrypted; the other information will remain encrypted. Here, the logged-on user has access to view the information of EmployeeID: 1, 2, and 3; but doesn t have enough privileges to view the information related to EmployeeID: 4,5,6,7.

229 208 Fig 12: Encryption of the Basic Salary column Fig 13: the encrypted excel file Convert the Decrypted XML file to Excel: The values of both InputXMLFile() and OutPutExcelFile() properties will be set in order to call the ConvertXMLtoExcel() method. The result will be posted on DataGridView3 B. Excel file supporting Macros: emp.xlsm in Visual Basic for Application. Fig 14: Displaying the EmployeeDecryptedExcel.xlsx file The Btn_ConvertDecrypted_XMLtoExcel_Click() method will perform the following actions: Create Command Buttons and set their properties: 1. Go to Developer tab>insert>activex Control> Command Buttons 2. Drag 5 command buttons and set their names respectively: ConvertXMLtoExcel, EncryptSelectedNodes, EncryptAll, Create_Encrypted_Excel, Decrypt and Create_Decrypted_Excel. 3. Set their captions respectively to: Convert this Excel to XML, Encrypt Selected Nodes, Encrypt All BASICSALARY Column, Create Encrypted Excel, Decrypt, and Create Decrypted Excel.

230 209 Fig 15: emp.xlsm excel file Fig 16: employee.xml file Fig 17: encryptedxml.xml file

231 210 Fig 18: EncryptedExcel.xlsx file. Fig 19: decryptedxml.xml file Setting privileges to Windows logged-on users: Logged-on users have different privileges between each others. To give a certain user a permission to view a portion of data while other logged-on user is given other portion, we have to add a meaningful description to them. This description may present their positions in the Company. This description is used in the decryption process and leads to different results. In Fig 20: decryptedexcel.xlsx file our case, we have two logged-on users: thawari and Rouwa. thawari is a Manager of the first branch of the EGC company and must grant access only to information relevant to EmployeeID: 1,2, and 3 while Rouwa is a Manager of the second branch of this company and must grant access only to information relevant to EmployeeID: 4,5,6 and 7. Adding Description to windows logged-on Users:

232 211 Fig21: Adding Description to user thawari Fig22: Adding Description to user Rouwa Fig23: Decryption File of user thawari Getting User Information using WMI: In order to get the description of the current user in the local PC, we must know some information concerning Windows Management Instrumentation. WMI: Windows Management Instrumentation Referring to [13], WMI is the instrumentation and plumbing through which almost all Windows resources can be accessed, configured, managed, and monitored. Fig24: Decryption file of user Rouwa There are three steps common to any WMI script used in the script to retrieve information about a WMI managed resources: connecting to WMI service, retrieving a WMI managed resource and displaying properties of WMI managed resource. Establishing a connection to the Windows Management Service on a local or remote computer is done by calling VBScript's Getobject function and passing GetObject the name of the WMI Scripting Library's moniker, which is "winmgmts:" followed by the name of the target computer.

233 212 WMI Architecture The WMI architecture consists of three primary layers: Managed resources, WMI infrastructure and Consumers. Classes are grouped into namespaces, which are logical groups of classes representing a specific area of management. For example, the root\cimv2 namespace includes most of the classes that represent resources commonly associated with a computer and operating system. WMI Providers WMI providers act as an intermediary between WMI and a managed resource. Providers request information from, and send instructions to WMI managed resources on behalf of consumer applications and scripts. For example, WIN32 provider provides information about the computer, disks, peripheral devices, files, folders, file systems, networking components, operating system, printers, processes, security, services, shares, SAM users and groups, and more. 4. Conclusions Fig25: WMI Infrastructure Managed Resources A managed resource is any logical or physical component, which is exposed and manageable by using WMI. Windows resources that can be managed using WMI include the computer system, disks, peripheral devices, event logs, files, folders, file systems etc A WMI managed resource communicates with WMI through a provider. WMI Infrastructure The middle layer is the WMI infrastructure. WMI consists of three primary components: the Common Information Model Object Manager (CIMOM), the Common Information Model (CIM) repository, and providers. Together, the three WMI components provide the infrastructure through which configuration and management data is defined, exposed, accessed, and retrieved. A fourth component, even if small, but absolutely essential to scripting is the WMI scripting library. CIM Repository WMI is based on the idea that configuration and management information from different sources can be uniformly represented with a schema. The CIM is the schema, also called the object repository or class store that models the managed environment and defines every piece of data exposed by WMI. The schema is based on the DMTF Common Information Model standard. Much like Active Directory's schema is built on the concept of a class, the CIM consists of classes. A class is a blueprint for a WMI manageable resource. To sum up, the idea of this paper has born to address a certain limitation in Windows Right Management Services which is the non-reusability of a file per different trusted recipients. Since WRMS provides a security on the whole document, it is not possible to share a part of the data in an RMS-enabled file with other trusted recipients having different privileges. Therefore, the sender is obliged to send to those recipients different files that do not contain confidential data, and applies WRMS on each file. This paper has introduced a new layer for securing sensitive data in Excel worksheet. It gives the flexibility to the sender to send only one copy of the file to different trusted recipients having different privileges and permissions. The provided solution will apply a security on a portion of the document which contains the confidential and sensitive data. As a final result, it does a selective access to the data: each receiver can only see data that grants access to. If this solution is accompanied with WRMS, the sender can get extra benefits such as: the Excel file can be kept secure toward unauthorized recipients in addition it can be ensured that the received data will not be changed. The advantages of this solution are: Efficiency in memory usage, time saving and file reusability in a proper secure way that fits many recipients having different privileges and protects the file from unintended or malicious recipients. The provided solution can be considered as scalable since the tremendous number of users does not affect on how the solution works. Since this solution was implemented as a separate module (DLL file), any update or improvement in

234 213 the code became easier to apply since it doesn t affect other parts of the program [14]. In addition, the separate module makes the solution user-friendly, since any basic programmer can get all its benefits simply by sending few parameters (such as name of certain files) without the need to master programming languages. Furthermore, this solution also provides a large compatibility and interoperability with different types of applications such as (Component Object Model) and different programming languages (such as C, Pascal, or standard call). This solution is not relying on WRMS; it is capable of working independently. In future works, this security layer should be applicable in all Microsoft Office applications (word documents, PowerPoint presentations, Microsoft Access ) and perhaps portable document file (.pdf) files. Furthermore, if the solution was extended to a web service, any application that supports web services can use it (such as: Java application, PHP application ). [9] XPath Tutorial: [10] Win32_userAccount Class: [11] Vicky Desjardins, Script to get local user, description, last logon, Group Membership for dummies. Published: Feb 15, 2007: [12] Impersonation of Client: ment.connectionoptions.impersonation.aspx [13] Windows Management Instrumentation: [14] Advantages of Dynamic Linking: References [1] Deb Shinder, How the Windows Rights Management Service can Enhance the Security of your Documents. Published: Sep 23, 2003 Updated: Apr 06, ghts_management_service_documents.html [2] Tony Bradley, NETWORK SECURITY TACTICS, Information protection: Using Windows Rights Management Services to secure data. Published: Aug 01, 2008: gci ,00.html [3] Technical Overview of Windows Rights Management Services for Windows Server 2003: [4] XrML, W3C. [5] Visual Basic.NET for first time programmers Workbook, Question Edition, Document Version 1.1 Copyright 2004 LearnVisualStudio.Net pages: 131, 132, and [6] Jaspal Singh, Excel Connectivity in VB.NET published: Aug 18, 2005: y.aspx [7] Import from excel 2007 into dataset problem: [8] Bilal Siddiqui, Exploring XML Encryption Part 1, demonstrating the secure exchange of structured data. Published: Mar 01, 2002:

235 An Enhanced Algorithm Of The Statistical Training Method In Boosting-Based Face Detection 214 Said Belkouch 1, Mounir Bahtat 1, Abdellah Ait Ouahman 1, M. M rabet Hassani 2 1 Micro-informatics, Embedded Systems and Systems On the Chips Lab., National School of Applied Sciences, University of Cadi Ayyad, Marrakech, Morocco 2 Faculty of Sciences Semlalia, University of Cadi Ayyad, Marrakech, Morocco Abstract A trained cascade for face detection with a reduced number of Haar-like features should be computationally efficient. The accurate classical scheme for selecting these Haar-like features is proposed by Viola and Jones, but the training process may take weeks. Recently, there have been several heuristics reducing the training time in a dramatic way but the selected weak classifiers are not as good as those chosen by Viola and Jones, which leads to an increased number of features in the final cascade and then decreasing detection speed. Our method is an improved version of a statistical training method; it presents both faster selections and accuracy comparable to the Viola and Jones method. Keywords face detection, Haar-like feature, weak and strong classifier, statistical training. 1. Introduction Face detection is a fundamental task for many applications such as face recognition, surveillance, smart homes and robotics. Several frameworks have been proposed to solve this open problem, from which, the Support Vector Machines (SVM) methods and especially the Viola and Jones algorithm [1] have gained a lot of interest as it can achieve very high detection rates in an extremely fast way over all existing efficient methods [2,3,4,5,6]. Viola and Jones framework is a cascade-based face detector, composed of a number of nodes where each node represents a classifier designed for a fast rejection of likely non-face sub-windows. These classifiers are constructed using a modified version of the AdaBoost algorithm [7] which is known to be very resistant to overfitting compared to other boosting methods [8]. A set of simple Haar-like features are used to train the classifiers. Many extensions of this set are proposed in [9, 10, 11] yielding to a trained cascade with a fewer number of features. In the present article we don t focus on which feature set is more convenient for training a face detector. The exhaustive search over the feature set that is proposed by Viola and Jones makes training a cascade of classifiers a time consuming part as the algorithm complexity in that case is O(NTlog(N) where N is the number of training images and T is the number of used features. This leads to a total time of weeks to train the full cascade. A first try to reduce this complexity was proposed in [12] reducing it to O(NT) using caching. This implementation is called Faster AdaBoost, but uses very huge memory space, and becomes harder to implement when the feature set is wider. The most interesting heuristic reducing this complexity is the one proposed in [13], which is a statistical method that succeeded to break up the NT factor to a complexity of O(Nd 2 +T), where d=24 2. In this case training images are scaled to a resolution of 24x24. This method is able to perform extremely faster training at the cost of slight decrease in accuracy. The statistical heuristic is not sensible to enlarging the feature set in terms of training time, which gives an opportunity to extend the feature set by more complicated ones. This was not possible using the classical method as the training time increases in an exponential way with the features number. This heuristic is one of the best known strategies to train a boosting-based cascade of classifiers taking into account the compromise of speed and accuracy. Our algorithm is an improvement over the last mentioned statistical heuristic. It is faster and more precise, and with the help of parameters adjustment our algorithm is able to choose exactly the same weak classifiers as those chosen by the precise Viola and Jones method. The key idea of our training method is a combination technique between the two discussed methods. Indeed, we observed that the estimated feature errors that were evaluated by the statistical algorithm could not be used as a decisive comparing tool between elements of the feature

236 215 set. For this reason, we exploited the statistical heuristic as a mean to candidate some good features and treat them just afterward in a more precise way using the classical algorithm of Viola and Jones. This increases the selection accuracy, and with some additional modifications that will be explained later we make our algorithm running faster. The remaining parts of this paper will be presented as follow: the Viola and Jones framework is exposed in section 2. In section 3 we will describe the fast selection of Haar-like features using the statistical method. Our method is presented in section 4. Associated implementation and results are presented in section 5 and finally the conclusion is given in section Training of a Strong Classifier Using the Viola and Jones Method The Viola and Jones architecture is based on a cascade of strong classifiers as illustrated in fig. 1, where each one is composed of a number of weak classifiers. Each weak classifier consists of three parameters: a Haar-like feature f from the set shown in fig. 2, a polarity p and a threshold θ. The weak classifier parameters as presented in equation (1) are chosen by the AdaBoost boosting algorithm. In this equation, h p,θ,f (x) is the weak classifier function and x is the 24 x24 input image. This function returns 1 to indicate that x is a face and 0 otherwise.,, 1 0 The feature value f(x) is computed by the sum of pixel intensities in the white region of the feature f applied on x subtracted from the sum of pixel intensities in the grey one. (1) Figure 2. Four types of Haar-like features, have led to different features The weak classifiers are expected to have big error rates, but as long we combine more weak classifiers into the cascade architecture, the global error rate tends to 0. The boosting algorithm requires a number of training images from two categories: face images and non-face images. In our experiments we have used 5000 face images from the LFW database [14] and a number of non-face images. AdaBoost serves to train a strong classifier for the cascade. It begins by initializing weights indicating the importance of each training image, and during each round of the algorithm a best weak classifier h t is selected. Finally, as a preparation phase for the next round, the weak classifier error ε t is computed and weights are updated so as the next weak classifier corrects the errors of the previous one. The final form of the trained strong classifier that contains S weak classifiers is given by equation (2): 1 0 (2) Where α ln The best weak classifier error is ε t ; evaluated as : ε ω h x y (3) Where y n = 1 when the x n image is a face and y n = 0 otherwise. Figure 1: A cascade of classifiers The process of selecting the best weak classifier h t is done by an exhaustive search over all possible weak classifiers, which it has a complexity of O(NTlog(N)) for the Viola Jones algorithm. This implies that for each feature, a quick sort of N feature values is needed.

237 216 The cascade is finally constructed by a combination of many strong classifiers with a well-chosen number of contained weak classifiers in the purpose of boosting detectors speed. 3. Fast Selection of Haar-Like Features Using the Statistical Method: The statistical method has improved the process of selecting the best weak classifier h t in terms of training time. It succeeded to break up the NT factor, by treating separately training samples and features. Instead of passing each feature on all the training images in order to determine its best parameters as it was the case in the Viola Jones algorithm, this method treats the N training samples only once per round. During this phase, statistics of these images are extracted and used later with each feature in order to compute the wanted parameters in a constant time. The best feature is then the one that best separates between feature values on face images and those on non-face images. The feature value associated with the m th feature is considered as a random variable that we will denote υ (m,c) where c = 1 when applied on the face class or c = 0 otherwise. The feature value consists of a linear expression of integral image values, then, if we denote y (c) the vectorized random variable form of the integral image from the c class, we will have:, T g Where g (m) is a vector describing the m th feature. The probabilistic law of the random variable υ (m,c) is estimated to be a Gaussian distribution over training images :, ~ µ,,σ, The mean parameter is given by: And the variance is : µ, g σ, g T δ g (4) (5) (6) (7) Where δ (c) is the covariance matrix of y (c) and is expressed by: δ T T The mean sign <.> is computed taking into account the current weighing distribution for training images; then we have: δ m y,, ω, ω,,t T m m The example on fig. 3 and fig. 4 shows the estimated probabilistic density function of υ (m,c) against the real one for a specified feature. We can see that the approximation is good enough regardless of a certain loss of precision. Let s consider the following notations: u ω, u ω, µ µ, µ µ, σ σ, σ σ, f µ, x 1 µ σ 2π e We suppose without loss of generality that: µ µ u u 1 For each feature, the optimum polarity ( p ), the threshold ( θ ) and the error ( ε ) are computed in constant time, in the following way: p signµ, µ, The θ parameter is the minimum of the error function εθ which is defined below: εθ u f µ, x dx u f µ, x dx (8) (9) (10) (11) (12) (13) (14) (15) (16) (17) (18) (19) (20)

238 217 in [13] that it takes about 1.8 seconds computation for that part using the highly optimized algebra package GotoBLAS [15]. As we don t have this package we have used in our implementation the classical matrix multiplication algorithm which is less efficient than the GotoBLAS implementation. Therefore it takes more time to compute that same part of the algorithm. 4. Our Improved Algorithm For Boosting- Based Training: Figure 3: The real density functions for υ (8447,c) The statistical training method is quite fast, but the process of weak classifiers selection terminates by choosing the wrong weak classifier due to the probability distribution estimation as shown in fig. 3 and fig. 4. As explained in the introduction, our method combines both the statistical and the Viola Jones training methods in a cascaded architecture as shown in fig. 5. The statistical part function aims to eliminate likely non-convenient features very quickly, leaving only a small amount of features that contains most probably the best weak classifier. This smaller set is analyzed in a more precise way using the Viola and Jones algorithm, which can be also done in a short time due to the very few elements in the resulting feature set. Figure 4. The estimated density functions for υ (8447,c) The best error is then given by: ε εθ (21) The statistical selection part will then be done as the following form: For each class c do Compute the vector Compute the symmetric matrix End for For each feature For each class c Compute µ, Compute σ, Compute, and Compare with the best End For End for This has improved the complexity to ; the bottleneck of this method is at computing the matrix which has a complexity of. It has been reported Figure 5: The architecture of our method The reduction percentage can be adjusted empirically; it controls a compromise between the accuracy and the speed. Then, according to our architecture, the training method is certainly going to be more accurate than the statistical method whatever could be. In order to select the best first % elements, a sort operation is needed for the whole set of features according to their errors, which has a complexity of ln; then the full complexity of the algorithm is 102. This complexity will take more time than the statistical method and should be the smallest possible in order to have a short total training time. This first implementation will be denoted.

239 218 The first filtering block in our architecture which is the statistical algorithm doesn t make a decisive choice about the best weak classifier. It is permitted that it makes errors on evaluating each feature performance as these errors are compensated by a good choice of the parameter. Then this architecture allows us to make some simplifications on the statistical algorithm in order to boost computation speed while a slightly modified could remove the engendered errors. Thus, we have decided to operate on rescaled training images from to during the first filtering bloc. The second bloc will still be operating on training images. This will have the effect to reducing the complexity term by a factor of 16. According to our experiments, if is well-chosen, then the term is much greater in the complexity than other terms; then reducing it will reduce the training time significantly. This implementation will be called and it is shown in fig. 6. implementations of our method 1, with the Viola Jones implementation and the statistical implementation. Fig. 9 compares the implementation in the same manner. These results are obtained using the feature set 1. The obtained results in fig. 8 show that the Viola and Jones method converges faster than the statistical method, which means it uses only a small amount of features to reach the same performance. Indeed for example, 2.02 % error rate is reached with 20 features for Viola and Jones method while it needs 50 features for the statistical method. Figure 6. The implementation 5. Implementation and Results For our experiments we have used for training 5000 face images from the cropped version of the LFW (Labeled Faces in the Wild) that are distributed by the university of Massachusetts [14], and non-face images taken from the web, all training images have been rescaled to a resolution. Figure 7.A part of the feature set 2 Two different sets of features have been used. The first set that we will denote 1 contains features using the 4 feature types presented in fig. 2. The second denoted 2 contains features using 92 feature types and that some of them are presented in fig. 7. We have developed a program that can generate the feature vectors g after drawing the feature form manually, so that increasing the number of features can be done easily. All experiments are done in a Core2 Quad 2.4 GHz CPU. Experimental results are presented in figs. 8 and 9. Performance is measured by comparing the evolution of the error rate in a strong classifier. Fig. 8 compares three Figure 8: implementation results

240 219 Our α implementation decreases significantly the error rate when compared to the statistical method, even when choosing 0.01 with which we have the fastest α implementation. In this case the full feature set is decreased by a factor of 10 (i.e. 0.01%) to the second part of our cascaded architecture. In fig. 9, the implementation has been used. It presents higher error rate than the α implementation, but still converging more quickly than the statistical method resulting in significant decrease in training time as shown in Table 1. The statistical method takes 31 seconds in our implementation, which is greater than that mentioned in [13], that s due as explained in section 3 to their use of a non-classical multiplication algorithm that s presented in the highly optimized algebra package GotoBLAS. If the algebra package had been used in our case, the α and β methods training time would further decrease as well. The α implementation takes slightly more training time than the statistical method as shown in table 1 but shows a huge increase in accuracy as illustrated in fig. 8. Furthermore, for k=1 our method presents almost the same accuracy as the Viola and Jones method by the approximately superposed curves while it is worth noticing that the training time takes 48 seconds in our case and 31 minutes with Viola and Jones Method. The implementation loses some accuracy comparatively to α implementation as showed in fig. 9, but decreases the training time significantly. We notice that the method has improved accuracy over the statistical method and furthermore decreases the training time as low as 2 seconds for k = To the best of our knowledge this is the fastest training method in boosting-based face detection using Haar-like features. An equivalent statistical method using a rescaling to would run as fast as our method, but would increase error rates to huge numbers according to our experiments. Table 1: Training time on the feature set (1) Method Statistical method Viola and Jones method Our α method k=1 Our α method k=0.2 Our α method k=0.01 Our method k=1 Our method k=0.2 Our method k=0.01 Training time of a weak classifier 31 seconds 31 minutes 48 seconds 34 seconds 31seconds 16 seconds 5 seconds 2 seconds Fig. 10 shows the effect of increasing the features number. The feature set (2) has been used with the α implementation of our method for k=1 and it is compared to the Viola and Jones method using the feature set (1). The figure shows that the error rate has been decreased in comparison with the Viola and Jones method. Besides, our method allows to increase the number of features in the training process without significant increase in the training time as shown in table 2. It is worth to remember that it does not make a sense to use Viola and Jones with feature set (2) because training then would not be praticly possible. Indeed as shown in table 2, training time of only one weak classifier takes 3.2 hours. Figure 9: implementation results

241 220 Our method also allows enlarging the feature set so as to attain better performance in terms of convergence over the training set. REFERENCES Figure 10. The effect of increasing the number of features using the implementation Table 2: Training time on the feature set (2) Method Statistical method Viola and Jones method Our α method k=1 Our α method k=0.2 Our α method k=0.01 Our method k=1 Our method k=0.2 Our method k= Conclusion: Training time of a weak classifier 50 seconds 3.2 hours 2.6 minutes 1.1 minutes 51 seconds 1.6 minutes 30 seconds 14 seconds In this paper we present a fast and accurate selection method of Haar-like features, two implementations have been proposed according to the same architecture. The α implementation presents a high accuracy that can be comparable to the Viola and Jones exhaustive search method while slightly increasing the training time relatively to the statistical method, the implementation loses some accuracy while still being more accurate than the statistical algorithm, besides, it decreases the training time over all known existing Haar-like selection strategies. [1] P. Viola and M. Jones, Rapid Object Detection Using a Boosted Cascade of Simple Features, IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), vol. 1, pp , 2001 [2] K. Sung and T. Poggio, Example-Based Learning For View-Based Face Detection, IEEE Patt.Anal. Mach. Intell., vol. 20, pp , 1998 [3] H. Rowley, S. Baluja, and T. Kanade, Neural Network-Based Face Detection, IEEE Patt.Anal. Mach. Intell,vol. 20, pp , 1998 [4] H. Schneiderman and T. Kanade, A Statistical Method For 3D Object Detection Applied To Faces And Cars, IEEE Conference on Computer Vision and Pattern Recognition. Proceedings , USA, vol. 1, pp , 2000 [5] M. Yang, D. Roth, and N. Ahuja. A Snowbased Face Detector. In NIPS-12; Conference on Advances in Neural Information Processing, Systems, pp. 855,-861. MIT Press, [6] Y. Amit, D. Geman, and K. Wilder, Joint induction of shape features and tree classifiers, IEEE Transaction on Pattern Analysis And Machine Intelligence, vol. 19, no. 11, pp , 1997 [7] Y. Freund and R. E. Schapire. Experiments With a New Boosting Algorithm. In Machine Learning, In Proceedings of the Thirteen International Conference In Machine Learning, Bari, pages , [8] P. L. Bartlett and M. Traskin, AdaBoost is consistent, Journal of Machine Learning Research, vol. 8, pp , [9] R. Lienhart and J. Maydt, An Extended Set of Haar-Like Features For Rapid Object Detection, IEEE 2002 International Conference on Image Processing, Vol. 1, pp , Sep [10] T. Mita, T. Kaneko, O. Hori, Joint Haar-Like Features For Face Detection, Tenth IEEE International Conference on Computer Vision, ICCV 2005, Beijing. Vol. 2, pp , 2005 [11] K. Masada, Q. Chen, H. Wu and T. Wada, GA Based Feature Generation For Training Cascade Object Detector, 19th International Conference, ICPR 2008, pp.1-4, Tampa FL, 2008 [12] J. Wu, S. C. Brubaker, M. D. Mullin, and J. M. Rehg. Fast asymmetric learning for cascade face detection. IEEE Trans Pattern Anal Mach Intell Mar;30(3): [13] M.T.Pham and T.J.Cham, Fast Training And Selection of Haar Features Using Statistics in Boosting-Based Face Detection, In Proc. International Conference on Computer Vision (ICCV 2007), Rio de Janeiro, Brazil, 2007 [14] Source Labeled Faces in the Wild Database, Computer Visions Lab., University of Massachussets, 2010, [15] K. Goto and R. van de Geijn, High-performance implementation of the level-3 Blas, FLAME Working Note #20. The University of Texas at Austin, Department of Computer Sciences. Technical Report TR Said Belkouch has completed his PhD in Microelectronics at University Joseph Fourier-Grenoble in France in From 1989 to 2003, he worked respectively as Assistant Researcher at University of Sherbrooke in Canada, Research Officer at National Council of Canada, and embedded ASICs Design Engineering at Tundra Semiconductor Corporation, Ottawa, Canada (Tundra acquired recently by Integrated Devices Technology Company). Since 2003, He is professor at Electrical Engineering Department, National School of Applied Sciences-Marrakech, Morocco. His area of research includes embedded systems and

242 221 microelectronics. He has published several research papers in Journals and Proceedings. Mounir Bahtat is a Master and Engineer student at Microinformatics, Embedded Systems and Systems On the Chips Lab Abdellah Ait Ouahman received the doctorate thesis in Signal Processing from the University of Grenoble, France, in November His research was in Signal Processing and Telecommunications. Then he received the PhD degree in Physics Science from the University of Sciences in Marrakech, Morocco, in He is now Professor and responsible of the Telecommunications and Computer Science and Networking laboratory in the Faculty of Sciences Semlalia in Marrakech, Morocco. His research interests include the signal and image processing and coding, telecommunications and networking. Actually he is a director of National School of Applied Sciences, Marrakech. He has published several research papers in Journals and Proceedings. Moha M Rabet Hassani. He received the a doctorate thesis in automatic from Nice University, France, in 1982 and Ph. D degree in electrical engineering from Sherbrook university, Canada, in He is now Professor in the Faculty of Sciences Semlalia, Cadi Ayyad University, Marrakech, Morocco. He heads at the University both the Doctoral Studies Centre of Science and Technology and Electronics and Instrumentation Lab. His research interests are in statistical signal processing, nonlinear system, identification and equalization fields.

243 222 Personnel Audit Using a Forensic Mining Technique Adesesan B. Adeyemo 1 and Oluwafemi Oriola 2 1 Computer Science, University of Ibadan Ibadan, Oyo State, Nigeria 2 Computer Science, University of Ibadan Ibadan, Oyo State, Nigeria Abstract This paper applies forensic data mining to determine the true status of employees and thereafter provide useful evidences for proper administration of administrative rules in a Typical Nigerian Teaching Service. The conventional technique of personnel audit was studied and a new technique for personnel audit was modeled using Artificial Neural Networks and Decision Tree algorithms. Atwolayer classifier architecture was modeled. The outcome of the experiment proved that Radial Basis Function Artificial Neural Network is better than Feed-forward Multilayer Perceptron in modeling of appointment and promotion audit in layer 1 while Logitboost Multiclass Alternating Decision Tree in Layer 2 is best in modeling suspicious appointment audit and abnormal promotion audit among the tested Decision Trees. The evidential rules derived from the decision trees for determining the suspicious appointment and abnormal promotion were also presented. Keywords: Data Mining, Forensic, Audit, Appointment, Promotion 1. Introduction By general definition, audit is the evaluation of a person, organization, system, process, enterprise, project or product. There are two types of audit namely: Financial Audit and Compliance Audit. In organizations, personnel are employed and assigned to various departments or units based on certain criteria which include skill, qualification, experience, and social/gender status. It is the desire and goal of every employer to secure and maintain a set of personnel that are legitimate within the consent/service condition of the institution and as such, verification and validation of employees is done severally through various means to determine the employment status of its staff. This is referred to as Personnel Audit. Personnel Audit is an example of Financial Audit. According to the Employment Equity, Planning and Policy Development Division of Personnel Policy Branch, Treasury Board in Canada, personnel audit is the systematic, independent review and appraisal of all departmental (personnel) operations, to determine the efficiency, effectiveness and economy of the departmental (personnel) management practices and controls. Personnel Audit entails verification and validation of the compliance of appointment, promotion, payment, background and job performance records. In this work, the focus will be on the audit of appointment and promotion records because of their significance and the challenges posed in Nigerian environment. Appointment Audit refers to a process that involves verification and validation of employees to ascertain their compliance with principles guiding recruitment and assignment to jobs. In this aspect, the employees can either be legitimate or illegitimate (ghost or ineligible). Legitimate employees refer to group of employees whose appointments comply with rules and regulations that govern its implementation within an organization. Ghost employees refer to group of employees that are not in existence and fraudulently enjoying staff benefits while Ineligible employees refer to group of employees that are not competent or fit for appointment. Promotion Audit is a process that involves verification and validation of employees to ascertain their compliance with principles guiding advancement and progress of employees through cadres or positions. In this aspect, the employee can be categorized as having a normal or abnormal promotion, being qualified for promotion or unqualified for a promotion. Abnormal promotion is a promotion that does not follow the due process and regulations. Employees qualified for promotion refers to employees whose promotion are indiscriminately impeded or delayed, while employees unqualified for promotion

244 223 refer to employees that occupy positions that they are not yet entitled to. Audit, in some cases can be Forensic in its approach depending on the level of duty or complexity. Forensic Science has been defined as the application of science and technology to the potential for evidence, at all stages of the investigative process, so that it can be located, recovered, analyzed and interpreted for the purpose of impacting on crime and criminality in a way that supports the effective administration of justice and inspires public confidence (Mennell, 2009). The primary objectives of forensic science include: Bringing offenders to justice; exonerating the innocent; detecting crime; ensuring efficient and effective investigations and gaining a better understanding of criminality, for example by understanding criminal behavior, links and associates via information provided through forensic data such as fingerprints, footwear marks, drug composition and tool marks. Forensic data mining which takes its roots from Forensic Science can be described as the application of data mining techniques and other scientific tools to investigative process for good and sound evidence (Chatterji, 2001). It often requires thorough IT knowledge of data matching and data mining techniques. This project intends to determine the authenticity of employees on the personnel and pay roll list of a Nigerian government agency using data mining techniques supported by forensic principles. 1.1 Forensic Data Mining Data Mining refers to the nontrivial extraction of implicit, previously unknown and potentially useful information from data in databases (Zaiane, 1999). It is a key step of knowledge discovery in databases (KDD). In other words, data mining involves the systematic analysis of large data sets using automated methods. By probing data in this manner, it is possible to prove or disprove existing hypotheses or ideas regarding data or information while discovering new or previously unknown information. It is noted for its Pattern Recognition ability that ensures that information is obtained from vague data (Baker, 2009). In particular, unique or valuable relationships between and within the data can be identified and used proactively to categorize or anticipate additional data. Through the use of exploratory graphics in combination with advanced statistics, machine learning tools, and artificial intelligence, critical nuggets of information can be mined from large repositories of data. While data mining has been applied to many areas of human endeavour such as business, education, manufacturing, and government, the application of data mining in personnel audit has not well been explored. In a report by Phua et al. (2004) on fraud detection based researches, it was shown that very few of these researches focused on employees fraud. Forensic Data Mining (or Forensic Mining) which originated from Forensic Science can be described as the application of data mining techniques and other scientific tools to investigative process for good and sound evidence. In the same vein, the investigative process of auditing called forensic auditing could be defined as the application of auditing skills to situations that have legal consequences (Chatterji, 2001). Forensic audit methodologies can be used to obtain a more detailed understanding of the entity and its activities to identify areas of risk both in determining the direction of the audit and in expressing an opinion. In forensic auditing, the following actions are important: working relations with the investigating and prosecuting agencies, authorisation and control of the audit investigation, documentation of relevant information and safeguarding all prime records pertaining to the case, rules of evidence governing admissibility or authentication of records, confidentiality, evaluation of the evidence to assess whether the case is sustainable, legal advice where appropriate and reporting the findings in a manner that meets legal requirements. As such, the knowledge of entity s business and legal environment, awareness of computer assisted audit procedures and innovative approach and sceptic of routine audit practices are required for forensic audit. The audit of personnel of personnel who are mostly teachers, laboratory attendants, administrative workers and guards in State s Teaching Service is carried out regularly to ensure that qualified employees and accurate record of personnel are on the state government payroll list. This is handled by internal auditors who are employees of the state civil service commission and external auditors who are consultants who work within the government auditing standard. However, the technique being used has not yielded reliable results. In our case study which is State Teaching Service, the conventional physical appearance screening of personnel and manual methods of personnel s records checking which are being used are faced with the problems of erroneous results, lack of support for effective administration of justice and inability in generating useful patterns that could be documented for later personnel audits in the organization. This poor result of audit of the State Teaching Service personnel has caused government a lot of resources in terms of training, payment of salary and other entitlements of illegitimate employees. Apart from this, a lot of wastage has been incurred, in terms of monetary waste and human waste during physical appearance screening of personnel due to poor method of processing and large volume of

245 224 data being processed. Therefore the techniques that are employed in carrying out the audit has posed many challenges like waste on the part of government and stakeholders in form of time, effort and money; and loss of lives and negative consequences on health of employees during audit exercise. In this research work, soft computing data mining methods (Artificial Neural Networks and Decision Trees) that are guided by forensic principles will be used to develop a model for personnel audit purpose using case data from one of the States Teaching Service Personnel Databases of in Nigeria. This will help to expose behavioural patterns of crime associated with appointment and promotion among employees in the Teaching Service, guide against human judgement that is distorted by an array of cognitive, perceptual and motivational biases applied during physical appearance screening of employees, reduce the anomalies in employees record, detect appointment and promotion fraud and give accurate staffing report in sufficient details to further allow accurate resizing and restructuring. 1.2 Decision Trees and Artificial Neural Networks A Decision Tree (DT) is a logical model represented as a binary or multiclass tree that shows how the value of a target variable can be predicted by using the values of a set of predictor variables. In the tree structures, leaves represent classifications and branches represent conjunctions of features that lead to those classifications. In decision analysis, a decision tree can be used visually and explicitly to represent decisions and decision making. The concept of information gain is used to decide the splitting value at an internal node. The splitting value that would provide the most information gain is chosen. Formally, information gain is defined by entropy. In other to improve the accuracy and generalization of classification and regression trees, various techniques were introduced like boosting and pruning. Boosting is a technique for improving the accuracy of a predictive function by applying the function repeatedly in a series and combining the output of each function with weighting so that the total error of the prediction is minimized or growing a number of independent trees in parallel and combine them after all the trees have been developed. Pruning is carried out on the tree to optimize the size of trees and thus reduce overfitting which is a problem in large, single-tree models where the model begins to fit noise in the data. When such a model is applied to data that was not used to build the model, the model will not be able to generalize. Many decision tree algorithms exist and these include: Alternating Decision Tree, Logitboost Alternating Decision Tree (LAD), C4.5 and Classification and Regression Tree (CART). An Artificial Neural Network (ANN) is a mathematical model or computational model that tries to simulate the structure and/or functional aspects of biological neural networks. It consists of an interconnected group of artificial neurons and processes information using a connectionist approach to computation. In most cases an ANN is an adaptive system that changes its structure based on external or internal information that flows through the network during the learning phase. Modern neural networks are non-linear statistical data modeling tools. They are usually used to model complex relationships between inputs and outputs or to find patterns in data. When creating a functional model of the biological neuron, there are three basic components of importance. First, the synapses of the neuron are modeled as weights. The strength of the connection between an input and a neuron is noted by the value of the weight. Negative weight values reflect inhibitory connections, while positive values designate excitatory connections. The next two components model the actual activity within the neuron cell. An adder sums up all the inputs modified by their respective weights. This activity is referred to as linear combination. Finally, an activation function controls the amplitude of the output of the neuron. An acceptable range of output is usually between 0 and 1, or -1 and 1. This process is illustrated diagrammatically in figure 1. From this model the interval activity of the neuron can be shown to be equal to an output function represented by Eq. (1) where the output of the neuron, y k, is the outcome of some activation function on the value of v k. Examples of ANN include Multilayer Perceptron (MLP) Neural Networks, Radial Basis Function (RBF) Neural Networks, Generalized FeedForward (GFFN) Neural Networks, Probabilistic and Generalized Regression Neural Networks. 2. Methodology 2.1 Classifier System Design Classifier Systems (Chan et al. (1999); Schetinin (2001) and Phua et al. (2004)) that combine artificial Neural Networks and Decision Trees were proposed.

246 225 The MLP and RBF Artificial Neural Networks were used for the modeling of the dataset in layer 1 while that the CART, Logitboost Alternating Decision and C4.5 decision tree algorithms were used for modeling the dataset layer 2. The comparison of the performance of the various algorithms was carried using standard metrics of accuracy, precision, recall and f-measure for classification. These are calculated using the predictive classification table called a Confusion Matrix (Table 1). Also Eq present the metrics used in the experiment and their definitions. Fig. 1: Artificial Neural Network Process Actual Table 1: Confusion Matrix Prediction A B TN FP a FP TP b In Mahesh et al. (2009), Decision Tree models recorded impressive prediction accuracy. On the other hand, the effectiveness of feed forward neural networks classifications has been tested empirically in (Guoqiang, 2000). Neural networks have been successfully applied to a variety of real-world classification tasks in industry, business and science (Widrow et al., 1994). In (Lampinen et al. (1998); Petche et al. (1998); Barlett et al. (1992) and Hoskins et al. (1990)), neural networks were applied to inspection of processes and detection and diagnosis of faults with good outcomes. Therefore in this compliance audit that has to do with inspection and detection, we experimented with a two layer classifier system for the proposed personnel audit system. Layer 1 consists of an Artificial Neural Network that models the appointment and promotion datasets, while Layer 2 is made up of a Decision Tree classifier that models suspicious appointment and abnormal promotion datasets derived from the output of the datasets of layer 1. This two-layer model reduces the size of the trees generated by the decision tree algorithm as the very effective and accurate neural networks would have been able to model the classification of the original dataset so that the decision tree classification of desired suspicious appointment and abnormal promotion dataset could yield a reliable result. Figure 2 presents the schematic diagram of the two layer Classifier System model. Different ANN and Decision Tree algorithms are tested on the dataset in order to determine that which best models the data at each layer. From Table 1: TN (True Negative) = Number of correct predictions that an instance is invalid FP (False Positive) = Number of incorrect predictions that an instance is valid FN (False Negative) = Number of incorrect predictions that an instance is invalid TP (True Positive) = Number of correct predictions that an instance is valid Accuracy = The proportion of the total number of predictions that were correct. Accuracy (%) = (TN + TP) / (TN + FN + FP + TP) (2) Precision = The proportion of the predicted valid instances that were correct: Precision (%) = TP / (FP + TP) (3) Recall = The proportion of the valid instances pages that were correctly identified Recall(%)= TP/(FN+TP) (4) F-Measure = This is derived from precision and recall values: F-Measure (%) = (2 x Recall x Precision) /(Recall+ Precision) (5) Sensitivity or true positive rate (TPR) equivalent to hit rate, recall TPR=TP/P=TP/(TP+FN) (6) Specificity (SPC) Or True Negative Rate SPC=TN/N=TN/(FP+TN)=1 FPR (7) The Kappa Statistics (κ): Is used to measure the concordance level between categorical data during prediction. Cohen's kappa measures the agreement between two raters that each classifies N items into C mutually exclusive categories. (8) where Pr(a) is the relative observed agreement among raters, and Pr(e) is the hypothetical probability of chance agreement, using the observed data to calculate the probabilities of each observer randomly saying each

247 226 category. If there is no agreement among the raters (other than what would be expected by chance), then κ 0. The F-Measure: Is used because despite the Precision and Recall values being valid metrics in their own right, one of them can be optimized at the expense of the other. The F- Measure only produces a high result when Precision and Recall are both balanced, thus this is very significant. The Receiver Operating Characteristic (ROC) curve: This shows the sensitivity (FN classifications) and specificity (FP classifications) of a test. The ROC curve is a comparison of two characteristics: TPR (true positive rate) and FPR (false positive rate). The TPR measures the number of valid instances that were correctly identified. TPR=TP/(TP+FN) (9) The FPR measures the number of incorrect classifications of valid instances out of all invalid test instances. FPR=FP/(FP+TN) (10) 2.2 Data Modeling and Pre-processing The data used for this research was collected from employee s data records through their credentials submitted for audit at the State s Auditor General s Office tagged as present data with suffix PR. Another set of data of the employees tagged as past data with suffix P containing original and duplicate information about employees was collected from the State Teaching Service Commission (employer). The data was cleaned, normalized and organized in a form suitable for data mining. WEKA version 3.6.2, an open source data mining software developed at the Waikato University was used for the data mining processing. Table 2 presents the attributes of appointment dataset, Table 3 presents the attributes of promotion dataset and Table 4 presents the categorization of data for the data mining process. The datasets were divided into two which includes the training and testing datasets. 66% of each of the datasets was devoted to training while the remaining 34% was used for testing of randomly selected new data. First MLP Neural Networks and RBF Neural Networks with varied parameters were used. The neurons in the hidden layers and the number of layers themselves were varied between 1 and 4 with a momentum value of 0.2, learning rate of 0.3, using the Gaussian activation function. These algorithms were used to model both appointment and promotion dataset. The Logitboost Multiclass Alternating Decision Tree (LAD), Classification and Regression Tree (CART) and C4.5 decision tree algorithms were used to model the suspicious appointment and abnormal promotion dataset. The experimental procedure used is: 1. Pre-processing: Load the formatted and normalized appointment dataset in the employees database (file) into the classifier application. 2. Set the number of hidden layers of MLP Artificial Neural Networks to zero, with momentum of 0.2 and Learning rate of Perform training and testing of dataset using percentage split option. If the result is satisfactory, stop modeling. 4. Repeat Experiment for one to four hidden layers of MLP Artificial Neural Networks if the result of step 3 is not satisfactory. 5. Repeat Experiment for one to four iterations of RBF Artificial Neural Networks. 6. Select the algorithm that best models the dataset based on the performance measures. 7. Load suspicious employees dataset drawn from appointment dataset into the application. 8. Perform training and testing of dataset using percentage split using LAD, CART and C Select the algorithm that best model the dataset based on the performance measures. 10. Repeat step 1 to 6 for promotion dataset in employees database. 11. Repeat step 7 to 9 for abnormal promotion dataset drawn from promotion dataset. Employee database input Decision Tree Classifier (Stage 2). Suspicious and abnormal records Figure 2: Neural Classifier Engine(Stage System Model 1) Fig. 2: Classifier System Model output Emplo yees status

248 227 Table 2: Attributes of Appointment Dataset Attribute Type Description Numerical Number on FormNo Verification Form SurnameP Categorical Past Surname SurnamePR Categorical Present Surname FirstNameP Categorical Past - First Name FirstNamePR Categorical Present - First Name PFileNoP Categorical Past - Personal File Number PFileNoPR Categorical Present- Personal File Number SexP Categorical Past Sex SexPR Categorical Present Sex HomeTownP Categorical Past- Hometown HomeTownPR Categorical Present-Home Town QualificationswithDateP Categorical Past - Qualification and Date of Qualification HomeTownP Categorical Past- Hometown HomeTownPR Categorical Present-Home Town QualificationswithDateP Categorical Past - Qualification and Date of Qualification QualificationswithDatePR Categorical Present- Qualification and Date of Qualification DateofBirthP Categorical Past- Date of Birth DateofBirthPR Categorical Present- Date of Birth DateofFirstAppP Categorical Past- Date on First Appointment DateofFirstAppP Categorical Present- Date of First Appointment GLon1stAppP Numerical Past- Grade level when first appointed GLon1stAppPR Numerical Present- Grade level when first appointed PresentGradeLevelP Numerical Past- Current Grade level PresentGradeLevelPR Numerical Present- Current Grade level StepP Numerical Past- Current Step StepPR Numerical Present- Current Step DatePostedPresentSchlP Categorical Past- Date Posted to Present School DatePostedPresentSchlPR Categorical Present- Date Posted to Present School NamePresentZEAP Categorical Past Name of Present Zonal Education Authority NamePresentZEAPR Categorical Present Name of Present Zonal Education Authority PresentRankP Categorical Past- Current Rank PresentRankPR Categorical Present Current Rank PresentSchlNameP Categorical Past Current School Name PresentSchlNamePR Categorical Present- Current School Name Status Categorical Appointment Status(Class) Table 3: Attributes of Promotion Dataset Attribute Type Description SurnameP Categorical Past Surname FirstNameP Categorical Past First Name PFileNoP Categorical Past- Personal File Number SexP Categorical Past- Sex QualificationwithdateP Categorical Past- Qualification and Date of Qualification DateofFirstAppP Categorical Past Date of First Appointment GLOn1stApptP Numerical Past Grade Level on First appointment DateofPreviousPromotionP Categorical Past- Date Promoted before the last promotion PreviousGradelevelP Numerical Past- Grade level in the previous promotion StepPreviousP Numerical Past Step in the previous promotion DateLastPromotionP Categorical Past- Date on the current promotion letter PresentGradelevelP Numerical Past- Grade level on the current promotion letter StepPresentP Numerical Past Current Step DatePostedPresentSchP Categorical Past Date Posted to Present School NamePresentZEAP Categorical Past Name of Present Zonal Education Authority PresentRankP Categorical Past Current Rank PresentSchlNameP Categorical Past Current School Name Status Categorical Promotion Status(class) Table 4: Categorization of data for the data mining process Number of Exemplars (Instances) Dataset Algorithm Training Testing Total Appointment ANN Suspicious Decision Appointment Tree Promotion ANN Abnormal Promotion Decision Tree RESULTS AND DISCUSSION The result of the Artificial Neural Networks Models of the Appointment dataset is presented in Table 5. The results show that RBF neural

249 228 network with two iterations was able to model the data better than the MLP neural networks. Also, the results of experiment on suspicious dataset using decision trees presented in Table 6 shows that the Logitboost Multiclass Alternating Decision Trees (LAD) was able to model the Suspicious Appointment Dataset better than the other algorithms used while the result of the promotion dataset and abnormal promotion dataset is presented in Table 7 and Table 8 respectively. The results of Table 7 show that RBF neural network with two iterations was able to model the data better than the MLP neural networks. Also, the results of experiment on abnormal promotion dataset using decision trees presented in Table 8 shows that the Logitboost Multiclass Alternating Decision Trees (LAD) was able to model the Suspicious Appointment Dataset better than the other algorithms used. The LAD tree models satisfied the support and confidence criteria that is greater than 0.5 (>0.5). In Table 9, the results of the weighted average measures of the selected best algorithms is presented with RBF recording TP rate, F-Measure and ROC that are above 0.8 while LAD tree recorded TP rate, F-Measure and ROC that are above 0.7 for all categories of dataset as presented in Table 4. Table 5: Results of Artificial Neural Networks Models of Appointment Dataset after Testing Performance Measure Instances Number of Layer (Multilayer Perceptron using Momentum= 0.2 and Learning Rate = 0.3) Number of Iteration (Radial Basis Function using Gaussian Radian Function) Time to model (Secs) % classification Correct Incorrect Kappa Statistic MAE RMSE RAE(%) RRSE(%) ROC Legitimate Suspect Precision Legitimate Suspect Recall Legitimate Suspect F-Measure Legitimate Suspect Table 6: Results of Decision Trees Models of Suspicious Appointment Dataset after Testing Performance Measure Instances LAD CART C4.5 Time to Model(seconds) Tree Size No of Leaves % Classification Correct Incorrect Kappa Statistics MAE RMSE RAE (%) RRSE (%) ROC Ghost Ineligible CannotSay Precision Ghost Ineligible CannotSay Recall Ghost Ineligible CannotSay F-Measure Ghost Ineligible CannotSay Table 7: Results of Artificial Neural Networks Models of Promotion Dataset after Testing

250 229 Performan ce Measure Time to model (secs) % classification Instances Multilayer Perceptron (Number of Layer) Momentum= 0.2 Learning Rate = 0.3 Radial Basis Function (Number of Iterations) Gaussian radian Function Correct Incorrect Kappa Statistic MAE RMSE RAE(%) RRSE(%) ROC Normal Abnormal Precision Normal Abnormal Recall Normal Abnormal F-Measure Normal Abnormal Table 8: Results of Decision Trees Models of Abnormal Promotion Dataset after Testing Performance Measure Instances LAD CART C4.5 Time to Model(seconds) Tree Size No of Leaves % Classification Correct Incorrect Kappa Statistics MAE RMSE RAE (%) RRSE (%) ROC Unqualified Referred Qualified Precision Unqualified Referred Qualified Recall Unqualified Referred Qualified F-Measure Unqualified Referred Qualified Table 9: Weighted Average Performance Measure of Best Algorithms RECORD ALGORITHM TP Rate FP Rate Precision Recall F-Measure ROC Area APPOINTMENT RBF SUSPICIOUS LAD APPOINTMENT PROMOTION RBF ABNORMAL LAD PROMOTION Table 10: Sample of the Staffing Report for Appointment Audit TR CannotSay TR CannotSay PFileNoP FormNo Rule Status TR CannotSay

251 230 NULL Ghost NULL Ghost NULL Ghost TR/2007/ ,12 Ineligible TR.2000/ ,12 Ineligible TR 2000/ Ineligible TR Comply Legitimate TR.2000/ Comply Legitimate TR 2000/ Comply Legitimate Table 11: Sample of the Staffing Report for Promotion Audit PFileNoP Rule Type NTS/PF/ Qualified TR/ Qualified TR/2001/01/192B 6 Qualified NTS/PF/ Referred NTS/PF/ Referred TR/14924/9 13,15 Referred NTS/PF/ Unqualified NTS/PF/ Unqualified Unqualified TR/2001/1466 Comply Normal TS/PF/4783 Comply Normal TR/37228/12 Comply Normal Table 12: Summary of Audit Report Audit Type Status Number of Employees Appointment Legitimate Comply 137 Suspect Ghost 47 Ineligible 12 Cannotsay 6 Promotion Normal Comply 102 Abnormal Unqualified 47 Referred 32 Qualified 30 In the Staffing Report of Table 10 and Table 11; and Summary of Results in Table 12, employees records with status Cannotsay and Referred means such records should be subjected to further probing as the data released and used for the audit cannot provide sufficient claims for true status of the employees. From the results obtained, the level of compliance of appointment records is above average (high significance) because the percentage compliance is more than 50% while the level of compliance of promotion records is below average (low significance) as the percentage compliance is less than 50%. 3.1 General Administrative Rules Rules generated from the LAD decision tree for suspicious appointment dataset are presented in Table 13 while rules generated from the LAD decision tree for abnormal promotion are presented in Table Conclusion The two-layer Classifier System model proposed and tested on case study data from a state Teaching Service Commission database was successfully used for Personnel Audit processing. The system which is based on a soft computing data mining process combines both ANN and Decision Tree algorithms in such a way that the ANN layer is first used to determine the compliance of records while the Decision Tree layer is used to determine derivation of behaviour patterns and rules from such records. The ANN layer uses a Radial Basis Function Neural Network while the Logitboost Multiclass Alternating Decision Tree algorithm was used in the Decision Tree layer. The two algorithms recorded an accuracy that was above 70% with an average F-Measure value of over 70%. This model is being further refined for the development of a Personnel Audit Expert System which will include identity verification using fingerprint and facial recognition based mining. Table 13: Administrative Rules Generated for Suspicious Appointments S/No Rule 1 IF (hometownpr = null) THEN Status = ineligible 2 IF(qualificationwithdatep=nce and hometownpr!=null) THEN status=ineligible 3 IF(presentgradelevelp<11 and qualificationwithdatep!=nce and hometownpr!=null) THEN status= ghost 4 IF(presentgradelevelp>=11 and qualificationwithdatep!=nce and hometownpr!=null) THEN status= Cannotsay 5 IF(hometownp= idanre and qualificationwithdate!= nce and hometownpr= null) THEN Status= cannotsay 6 IF(surname=ajelabi and presentgradelevelpr<15.5 and hometown!=idanre and qualificationwithdate!=nce and hometownpr!=null) THEN status = cannotsay 7 IF(firstnamep=Dickson and surname!=ajelabi and presentgradelevelpr<15.5 and qualificationwithdatep!=nce and hometownp!= null) THEN status= cannotsay 8 IF(surname=ajayi and firstnamep!= Dickson and surname!= ajelabi and presentgradelevelpr<15.5 and hometownp!= idanre and qualificationwithdatep!=nce and hometownpr!=null) THEN status=cannotsay 9 IF(surname!=ajayi and firstnamep!= Dickson and surname!= ajelabi and presentgradelevelpr<15.5 and hometownp!= idanre and qualificationwithdatep!=nce and hometownpr!=null) THEN status= ghost 10 IF(presentgradelevelpr>=15.5 and hometown!= idanreand qualificationwithdatep!=nce and hometownpr!=null) THEN status=cannotsay 11 IF(hometownp=idanre)THEN status=cannotsay 12 IF(hometownp!=idanre)THEN status= ineligible Table 14: Administrative Rules Generated for Abnormal Promotion S/No Rule 1 IF(presentrankp=typist and presentgradelevelp<13.5) THEN

252 231 status=unqualified 2 IF(qualificationwithdatep=nce91 and presentrank!=typist and prsentgradelevelp<13.5)then status=unqualified 3 IF(qualificationwithdate!=nce91 and present!=typist and presentgradelevelp<13.5) THEN status= qualified. 4 IF (presentgradelevelp<4.5 and previousgradelevel!=null and presentgradelevelp<13.5) THEN status=referred 5 IF(presentgradelevelp>=4.5 and previousgradelevel!=null and presentgradelevelp<13.5)then status=qualified 6 IF(presentgradelevelp<11 and previiousgradelevel!=null and presentgradelevelp<13.5)then status=unqualified 7 IF (presentgradelevelp>=11 and previousgradelevel!=null and presentgradelevelp<13.5) THEN status=referred 8 IF (presentgradelevelp>=11 and previousgradelevel!=null and presentgradelevelp<13.5) THEN status=referred 9 IF (presentgradelevelp>=11 and previousgradelevel!=null and presentgradelevelp<13.5) THEN staus=referred 10 IF(dateoffirstapptp!=28/10/1985 and presentgradelevelp>13.5) THEN status=unqualified 11 IF(stepprevious>=2.5 and presentgradelevelp>=13.5)then status=referred 12 IF(dateofpreviouspromotion=01/01/2003)THEN status=qualified 13 IF(dateofpreviouspromotion!=01/01/2003)THEN status= referred 14 IF(datepostedpresentschoolp=19/09/2008)THEN status=qualified 15 IF(datepostedpresentschoolp!=19/09/2008)THEN status=referred References [1] M.J. Abdolmohammedi, S.L. Elliot, Audit: West s Encyclopaedia of American Law, Full Article, Retrieved March3, 2010 from [2] B. Baker, Forensic Audit and Automated Oversight, Office of Auditor General based on logistic model tree. JBiSE. Vol.2, No.6, 2009, pp [3] E. B. Barlett and R. E. Uhrig Nuclear power plant status diagnostics using artificial neural networks, Nucl. Technol., Vol. 97, 1992, pp [4] P. Chan, W. Fan, A. Prodromidis and S. Stolfo Distributed Data Mining in Credit Card Fraud Detection, IEEE Intelligent Systems, Vol. 14: 1999, pp [5] A.N. Chatterji, Forensic Auditing. Conference of Common Wealth Auditor General, SAI-India. Retrieved April 3, 2010 from asosai_old/ journal2001 /style.css, [6] T. Cutting, How to Survive an Audit, PM Hut. Retrieved December 13, from [7] P. Z. Guoqiang, Neural Networks for Classification, in A survey. IEEE transactions on Systems, Man, and Cybernetics Part C: Applications and Reviews, 2000, Vol. 30, No. 4. [8] J.C. Hoskins, K. M. Kaliyur and D.M. Himmelblau D. M., Incipient fault detection and diagnosis using artificial neural networks, in Proc. Int. Joint Conf. Neural Networks, 1990, pp [8] J. Lampinen, S. Smolander and M. Korhonen, Wood Surface Inspection System based on Generic Visual Features, in Industrial Applications of Neural Networks,1998, pp [9] T. Petsche, A. Marcantonio, C.Darken, S.J. Hanson, G.M. Huhn, and I. Santoso, An autoassociator for on-line motor monitoring, in Industrial Applications of Neural Networks, 1998, pp [10] V. Mahesh, A. Kandaswamy, C. Vimal and B. Sathish ECG arrhythmia classification based on logistic model tree, JBiSE, 2009, Vol.2, No.6, pp [11] C. Phua, D. Alahakoon and V. Lee, Monitoring Report in Fraud Detection: Classification of Skewed Data, SIGKDD Exporation, 2004,Vol. 6, No 1, pp [12] C. Phua, V. Lee, K. Smith and R. Gayler, A Comprehensive Survey of Data Mining-based Fraud Detection Research, Retrieved March 17, 2010fromhttp:// m /fraud-detection-survey.pdf, [13] V.Schetinin, A Neural Network DecisionTreefor Learning Concepts from EEG Data, in NIMIA - SC NATO Advanced Study Institute on Neural Networks for Instrumentation, Measurement, and Related Industrial Applications: Study Cases Crema, Italy, 2001, pp [14] B. Widrow, D. E. Rumelhard, and M.A. Lehr, Neural networks: Applications in industry, business and science, in Commun. ACM, 1994, Vol. 37, pp [15] O. R. Zaiane, Principle of Knowledge Discovery in Databases, University of Alberta. Department of Computer Science. CMPUT First Author Adesesan B. Adeyemo obtained M.Sc and Ph. D. in Computer Science from Federal University of Technology, Akure, Nigeria in 1998 and 2008 respectively. He is a lecturer in Department of Computer Science of University of Ibadan, Nigeria. His research interest includes Data Mining and Computer Netwoks. He is a member of CPN and NCS in Nigeria Second Author Oluwafemi Oriola obtained M.Sc. in Computer Science from University of Ibadan, Ibadan, Nigeria in He works with SITA, Akure, Nigeria. His research interest includes Data Mining and Image Processing.

253 232 A RSS Based Adaptive Hand-Off Management Scheme In Heterogeneous Networks Debabrata Sarddar 1, Shovan Maity 1, Arnab Raha 1, Ramesh Jana 1, Utpal Biswas 2, M.K. Naskar 1 1. Department of Electronics and Telecommunication Engg, Jadavpur University, Kolkata Department of Computer Science and Engg, University of Kalyani, Nadia, West Bengal, Pin Abstract Mobility management, integration and interworking of existing wireless systems are important factors to obtain seamless roaming and services continuity in Next Generation Wireless Systems (NGWS).So it is important to have a handoff scheme that takes into account the heterogeneity of the network. In this work we propose a handoff scheme which takes handoff decision adaptively based on the type of network it presently resides and the one it is attempting handoff with through some predefined rules. It also relies on the speed of the mobile terminal to make a decision of the handoff initiation received signal strength (RSS) threshold value. Finally simulations have been done to show the importance of taking these factors into account for handoff decisions rather than having a fixed threshold value of handoff for different scenarios. Keywords: Received Signal Strength(RSS). Next Generation Wireless Networks (NGWS), Heterogeneous Wireless Networks, Quality of service(qos), Wireless LAN(WLAN), HIPERLAN. 1. Introduction With the rapid development of wireless technologies, the wireless networks have become more and more popular. Rapid research and development has led to the creation of different types of networks like Bluetooth, IEEE based WLAN, Universal mobile telecommunications system (UMTS) and satellite networks. These networks can be integrated to form Next Generation Wireless Systems that always provides the best possible features of different networks to provide ubiquitous connectivity. By connecting to any wireless access network, users can get many kinds of internet services out of doors. In wireless networks, mobility management provides mobile users to continuously get the internet service when they move between different subnets based on their service needs. With this heterogeneity, users will be able to choose radio access technology that offers higher quality, data speed and mobility which is best suited to the required multimedia applications. It is quite obvious from the discussion that it is an important and challenging issue to support seamless mobility and also QoS in order to support global roaming of mobile nodes (MN) in the NGWS. Handoff management is one of the most important features of mobility management. It is the process by which users keep their connections active when they move from one base station (BS) to another. In NGWS, two types of handoff scenarios may arise, horizontal handoff and vertical handoff [1], [2]. Horizontal handoff is defined as handoff between two BSs of the same network i.e. handoff between homogenous networks where one type of network is considered. Vertical handoff takes place between two BSs that belong to two different networks i.e. between heterogeneous networks and, hence, to two different Gateway Foreign Agents. The large value of signaling delay associated with the intra and intersystem handoff calls for the need of efficient threshold value of proper parameters for the support of delay-sensitive and real-time services. The delay in handoff is due to the several processes that have to take place for the handover of a MT from one BS to another. The handoff process involves two steps : Discovery and Reauthentication. The two steps mentioned above for a successful handoff introduce latency issues. These issues are as follows: Probe Delay This is the amount of time it takes the client to complete a scan of available networks and to build its priority list. Authentication Delay This is the amount of time it takes for the client to re-authenticate to the AP it chose from its priority list following any one of the different algorithms available. Re-association Delay This is the amount of time it takes for the client to signal the AP that the handoff is complete. There are five main handoff initiation techniques mentioned in [3], [4]: Received signal strength, relative signal strength with threshold, relative signal strength with hysteresis, and relative signal strength with hysteresis and threshold, signal to interference ratio based handoff. In our paper we

254 233 will mainly focus on the received signal strength based handoff techniques for handoff initialization. In received signal strength, the RSSs of different BSs are measured and the BS with strongest signal is chosen for handoff. In this paper a heterogeneous wireless access environment consisting of AP(WLAN,HIPERLAN), BS(Cellular Network) is considered as shown in Figure 1. A mobile node with multiple transceivers can get connected to these networks simultaneously. handoff algorithms that use received signal strength (RSS) information to reduce handoff latency and handoff failure probability are proposed in [6], [7], [8]. However, these algorithms are limited to handoff between third-generation (3G) cellular networks and WLANs, and do not take into account handoff between different networks. A novel mobility management system is proposed in [9] for vertical handoff between WWAN and WLAN. To achieve seamless mobility across various access technologies and networks, an MN needs information about the wireless network to which it could attach. Also, it is necessary to transfer information related to the MT from the current attachment point to the next one. The Candidate Access Router Discovery (CARD) protocol [10] and the Context Transfer Protocol (CXTP) [11] have been proposed to enable this procedure. Their key objectives consist of reducing latency and packet loss, and avoiding the re-initiation of signalling to and from an MT from the beginning. 3. Proposed Work: Fig 1: A Heterogeneous Network consisting of Cellular Network and WLAN and HIPERLAN We consider a geographic area which is totally covered by CN and is partially covered by WLAN and HIPERLAN Access Points.CN and WLAN, HIPERLAN are complementary to each other and hence we focus on the handoff scenario between these networks. A MT can be in the coverage area of a cellular network BS at one instant and be connected to the corresponding BS, but due to its mobility it can move over to the coverage area of a WLAN or HIPERLAN AP which lies within the CN coverage area. in some cases the coverage area of BS of CN can also overlap. Thus it is important that at any point of time the MT is connected to the proper attachment point (BS or AP) for service continuity, to enhance QoS factor of the network and also to keep the network congestion free. So it is important to have a proper handoff scheme between these networks. So it is quite obvious that multiple networks are involved I vertical handoff scheme. 2. Related Work: Work has been done to integrate WLAN/HIPERLAN/Cellular Network. Most of the work done is on architectures and mechanisms to support seamless mobility, roaming and vertical handoff. A vertical handoff decision method that simply estimates the service quality of available networks and selects the network with best quality is proposed in [5]. Different In this section we want to find the different received signal strength threshold value for handoff and also a proper scheme of handoff between neighboring cells in the Cellular Network (CN), between cellular Wireless Local Area Network(WLAN) and cellular network and viceversa, between High Performance Radio LAN (HIPERLAN) and cellular network and vice-versa. Received signal strength is a measure of the power present in a received radio signal. It determines the connectivity between a Mobile Terminal (MT) and Base Station(BS) or Access Point(AP).The Received Signal Strength(RSS) should be strong enough between BS/ AP and MT to maintain proper signal quality at the receiver. RSS gets weaker as a MT moves away from a BS/AP and the opposite happens when the MT moves closer to the BS/AP. So as MT goes away from the current BS/AP it is connected to handoff becomes necessary with its neighboring BS of CN or AP of WLAN. The RSS threshold value for handoff between different networks will be calculated in this section using formula of RSS for different networks. The threshold value of RSS depends on a few factors: 1. The velocity of the MT. 2. The latency of the handoff process. 3. The type of network the MT is presently in and the type of network with which the MT is trying to initiate handoff. 4. The size of the CN/WLAN/HIPERLAN cell the MT is presently residing. If the same threshold value of RSS is used irrespective of the handoff scenario then that will increase the probability of false handoff initiation which increases unwanted traffic resulting in the blocking of other calls. Also it will

255 234 increase the probability of handoff failure resulting in dropping of ongoing calls. So a different threshold value of RSS is used depending on the scenario of handoff. CELLULAR NETWORK Let r1 be the distance of the BS from the cell boundary of the cell the MT is presently situated.let x be the distance of the MT from the cell boundary of the present BS. Here hexagonal cells are considered. So as the MT moves towards the cell boundary RSS decreases from the first BS and it increases if the MT moves closer to the primary BS. We want to use RSS value to define a threshold, so that when the RSS drops below this threshold value handoff is initiated with the neighboring BS. MICROCELL The path loss in db for cellular network in micro cellular environment is given by PL= *log(f)-4.99*log(hbs) + [ *log (hbs)]*log (d) Here f=frequency in MHz d =distance in kilometres hbs =effective base station antenna height in meters The received signal strength for cellular network is expressed in dbm as Pcn =Pt + Gt -PL-A Here Pcn =received signal strength in CN Pt=transmitted power in db Gt=transmitted antenna gain in db A=connector and cable loss in db Now if the MT is at a distance x from the boundary of a WLAN cell whose size is a then we get that d= (.866*a-x). So we get that the path loss is given by PL= *log (f)-4.99*log (hbs) + [ *log (hbs)]*log (.866*a-x) If the MT is having a velocity v and the latency of handoff is T then for handoff failure probability to be zero the time taken by the MT to reach the boundary of the cell from the initiation of handoff t>=t. We take t=t to fulfill both the criteria of zero handoff failure probability and minimum false handoff initiation probability. So x=v*t So the received signal strength threshold is given by Pcnth=Pt+Gt-[ *log (f)-4.99*log (hbs) + [ *log (hbs)]*log (.866*a-(v*T))]-A If received signal strength decreases beyond this threshold value then the MT initiates handoff with its neighbouring BS or AP. WIRELESS LOCAL AREA NETWORK (WLAN) Log linear path loss model is given by PL=L+10*n*log(d)+S Here L=Constant power loss n=path loss exponent (values range between 2-4) d=distance between the MT and the WLAN AP in meters S=Shadow fading which is modeled with mean m=0 and standard deviation σ with values between 6-12 db depending on the environment. PL=path loss in db Now if the MT is at a distance x from the boundary of a WLAN cell whose size is a then we get that d= (.866*a-x) and x=v*t by similar arguments given in case of cellular networks. The received signal strength for a WLAN is expressed in dbm as Pw=Pt-PL Here Pw =RSS of WLAN in dbm The threshold value of signal strength for WLAN is given by Pwth=Pt-[ L+10*n*log(.866*a-(v*T))+S If the RSS is below this threshold value then the MT will initiate a handoff with the neighboring WLAN AP or the Cellular network BS. HIGH PERFORMANCE RADIO LAN (HIPERLAN) The propagation model for HIPERLAN considers geographic data (terrain, building, foliage and ground) to calculate the power in radio channel. Path loss indoor propagation model with shadow fading is given by PL= *log (d) +S Path loss outdoor propagation model with shadow fading is given by PL= *log (d) +.3* d+ S Here d=distance between the mobile terminal and the AP S=log normal shadowing its standard deviation=7 db for indoor and 8 db for outdoor PL=path loss in db

256 235 By the same argument as in case of WLAN if the MT is at a distance of x from the boundary of a cell of size a then d= (.866*a-x) and x=v*t. Hence PL= *log (.866*a-(v*T)) +S (For indoor propagation model) PL= *log (.866*a-(v*T)) +.3* (.866*a- (v*t)) + S (For outdoor propagation model) The received signal strength for HIPERLAN is expressed as PHL=Pt-PL Here PHL=Received signal strength of HIPERLAN in dbm Pt=Transmitted power in dbm. So received signal threshold value for indoor propagation model is given by PHLinth=Pt-[ *log (.866*a-(v*T)) +S] For outdoor propagation model this threshold value is given by PHLoutth=Pt-[ *log(.866*a-(v*T)) +.3* (.866*a-(v*T)) + S] In this case also handoff is initiated after RSS decreases beyond this threshold value. HANDOFF SCHEME In our proposed scheme the MT will not start handoff execution every time after RSS has fallen below a certain threshold level. We define certain rules for handoff to take place. Let RSS1= The RSS of the MT from the BS or AP it is presently connected. RSS2=the RSS from the BS or AP the MT is attempting handoff. RSSth(m,n)=RSS threshold value for handoff from network m to network n. Our proposed handoff execution scheme is as follows: i)when the MT is currently connected to a Cellular Network BS and is attempting handoff with a WLAN or HIPERLAN AP. Then handoff initiation will take place only when RSS2 >RSSth(n,m) m=cellular Network n=wlan/hiperlan Handoff in this direction only consider the RSS of the network with which it is attempting handoff and not the RSS of the present network because the CN is the largest network in our case and WLAN / HIPERLAN are within the CN. Handoff of this kind takes place to decrease the congestion in the larger Cellular Network, which handles heavy traffic. ii)when the MT is currently connected to a WLAN or HIPERLAN AP and is attempting handoff with a cellular network BS or another HIPERLAN/WLAN AP, then handoff initiation will take place when RSS1<RSSth(m,n) and RSS2>RSSth(n,m) m= WLAN/HIPERLAN n= Cellular Network/WLAN/HIPERLAN Handoff in this direction consider both the RSS of the present network and the network it is attempting handoff with, because if the handoff does not take place properly then the connection will be lost and hence handoff in this direction has more priority. iii) When the MT is currently in a Cellular Network and is attempting handoff with another BS of the Cellular Network,then handoff initiation takes place when RSS1<RSSth(m,n) m=n=cellular Network 4. Simulation Results: First we see that for a microcellular network the magnitude of received signal strength threshold decreases as the velocity of the MT increases for same value of handoff latency. This clearly shows that the RSS threshold should be dependent on the velocity of the MT. The simulation below shows the dependence of RSS threshold on the velocity of MT for different values of handoff latency. RSS THRESOLD VS VELOCITY OF MT IN MICROCELLULAR NETWORK (DIFFERENT LATENCY) It can be also seen that the RSS threshold value differs based on the size of the cell the MT is currently residing for same value of velocity. The simulation below shows that for smaller cells the magnitude of RSS threshold value

257 236 is less that is the handoff is to be initiated closer to the BS. RSS THRESOLD VS VELOCITY OF MT IN MICROCELLULAR NETWORK (DIFFERENT CELLSIZE) There is a strong dependence of RSS threshold value for handoff on the velocity of the MT for other kinds of networks also. But the dependence is different for different networks and also the threshold value is different for same handoff latency in different kinds of network. This is shown in the simulation results given below. RSS THRESOLD VS VELOCITY OF MT IN HIPERLAN (INDOOR) NETWORK (DIFFERENT LATENCY) RSS THRESOLD VS VELOCITY OF MT IN HIPERLAN (OUTDOOR) NETWORK (DIFFERENT LATENCY) RSS THRESOLD VS VELOCITY OF MT IN WLAN NETWORK (DIFFERENT LATENCY) All these results stresses the fact that having a fixed threshold value of RSS for different scenario will severely hamper the performance of the network, increasing congestion, reducing the QoS.

258 Conclusion : In this paper we explore the different handoff scenarios that can take place in NGWS. We have also proposed a scheme in which the handoff decision will depend on the type of network the MT is presently in and also the type of network it is attempting handoff to ensure least amount of handoff failure probability, thus providing sufficient QoS for delay sensitive and real time services. Effective handoff schemes also ensure minimal false handoff initiation probability, which leads to congestion and hence dropping of calls. Our simulation results shows how the received signal strength threshold for handoff varies according to different networks for the same value of MT velocity. This is the main point we want to convey in our paper. 6.REFERENCES: [1] I.F. Akyildiz, J. Xie, and S. Mohanty, A Survey on Mobility Management in Next Generation All-IP Based Wireless Systems, IEEE Wireless Comm., vol. 11, no. 4, pp , Aug [2] M. Stemm and R.H. Katz, Vertical Handoffs in Wireless Overlay Networks, ACM/Springer J. Mobile Networks and Applications (MONET), vol. 3, no. 4, pp , [3] Gregory P. Pollioni, Trends in Handover Design, IEEE Communications Magazine, vol. 34, March 1996, pp [4] P. Marichamy, S. Chakrabati and S. L. Maskara, Overview of handoff schemes in cellular mobile networks and their comparative performance evaluation, IEEE VTC 99,vol. 3, 1999, pp [5] N.Nasser,A. Hasswa and H.Hassanein, Handoffs in 4G Heterogeneous Networks,IEEE Communications Magazine,vol 44,no.10 pp oct [6] M.M. Buddhikot, G. Chandranmenon, S. Han, Y. Lee, S. Miller, and L. Salgarelli, Design and Implementation of a WLAN/ CDMA2000 Interworking Architecture, IEEE Comm. Magazine, vol. 41, no. 11, pp , Nov [7] H. Yokota, A. Idoue, T. Hasegawa, and T. Kato, Link Layer Assisted Mobile IP Fast Handoff Method over Wireless LAN Networks, Proc. ACM MOBICOM 02, pp , Sept [8] Q. Zhang et al., Efficient Mobility Management for Vertical Handoff between WWAN and WLAN, IEEE Comm. Magazine, vol. 41, no. 11, pp , Nov [9] Qian Zhang, Chuanxiong Guo, Zihua Guo, and Wenwu Zhu, Wireless and Networking Group, Microsoft Research Asia «Efficient Mobility Management for Vertical Handoff between WWAN and WLAN IEEE, communications Magazine, 2003 [10] M. Leibsch, A. Singh, H. Chaskar, D. Funato, E. Shim, Candidate Access Router Discovery (CARD), IETF RFC 4066, July [11] J. Loughney, M. Nakhjiri, C. Perkins, R. Koodli, Context Transfer Protocol (CXTP), IETF RFC 4067, July Author Biographies: Debabrata Sarddar is currently pursuing his PhD at Jadavpur University. He completed his M.Tech in Computer Science & Engineering from DAVV, Indore in 2006, and his B.E. in Computer Science & Engineering from Regional Engineering College, Durgapur in He was earlier a lecturer at Kalyani University. His research interest includes wireless and mobile system. Shovan Maity is presently pursuing B.E. (3rd Year) in Electronics and Telecommunication Engg. at Jadavpur University. His research interest includes wireless sensor networks and wireless communication systems. Arnab Raha is presently pursuing B.E. (3rd Year) in Electronics and Telecommunication Engg. at Jadavpur University. His research interest includes wireless sensor networks, advanced embedded systems and wireless communication systems. Ramesh Jana is presently pursuing M.Tech (2nd Year) in Electronics and Telecommunication Engg. at Jadavpur University. His research interest includes wireless sensor networks, fuzzy logic and wireless communication systems. Utpal Biswas received his B.E, M.E and PhD degrees in Computer Science and Engineering from Jadavpur University, India in 1993, 2001 and 2008 respectively. He served as a faculty member in NIT, Durgapur, India in the department of Computer Science and Engineering from 1994 to Currently, he is working as an associate professor in the department of Computer Science and Engineering, University of Kalyani, West Bengal, India. He is a coauthor of about 35 research articles in different journals, book chapters and conferences. His research interests include optical communication, ad-hoc and mobile communication, semantic web services, E-governance etc.

259 238 Mrinal Kanti Naskar received his B.Tech. (Hons) and M.Tech degrees from E&ECE Department, IIT Kharagpur, India in 1987 and 1989 respectively and Ph.D. from Jadavpur University, India in He served as a faculty member in NIT, Jamshedpur and NIT, Durgapur during and respectively. Currently, he is a professor in the Department of Electronics and Tele-Communication Engineering, Jadavpur University, Kolkata, India where he is in charge of the Advanced Digital and Embedded Systems Lab. His research interests include ad-hoc networks, optical networks, wireless sensor networks, wireless and mobile networks and embedded systems. He is an author/co-author of the several published/accepted articles in WDM optical networking field that include Adaptive Dynamic Wavelength Routing for WDM Optical Networks [WOCN,2006], A Heuristic Solution to SADM minimization for Static Traffic Grooming in WDM unidirectional Ring Networks [Photonic Network Communication, 2006], Genetic Evolutionary Approach for Static Traffic Grooming to SONET over WDM Optical Networks [Computer Communication, Elsevier, 2007], and Genetic Evolutionary Algorithm for Optimal Allocation of Wavelength Converters in WDM Optical Networks [Photonic Network Communications,2008].

260 239 A survey of Named Entity Recognition in English and other Indian Languages Darvinder kaur 1, Vishal Gupta 2 1 Department of Computer Science & Engineering, Panjab university Chandigarh , India 2 Department of Computer Science & Engineering, Panjab university Chandigarh , India Abstract In this paper, a survey is done on various approaches used to recognize name entity in various Indian languages. Firstly, the introduction is given about the work done in the NER task. Then a survey is given about the work done in recognition of name entities in English and other foreign languages like Spanish, Chinese etc. In English language, lots of work has been done in this field, where capitalization is a major clue for making rules. Secondly, a survey is given regarding the work done in Indian Languages. As Punjabi is one of the Indian languages and also the official language of Punjab. In next part, survey is given on Punjabi Language regarding what work is done and what work is going on in this field. Keywords: Named Entity, Named Entity Recognition, Tag set. 1. Introduction The term Named Entity, the word Named restricts the task to those entities for which one or many rigid designators stands as referent[22]. It is widely used in Natural Language Processing (NLP). It is the subtask of Information Extraction (IE) where structured text is extracted from unstructured text, such as newspaper articles. The task of Named Entity Recognition is to categorize all proper nouns in a document into predefined classes like person, organization, location, etc. NER has many applications in NLP like machine translation, question-answering systems, indexing for information retrieval, data classification and automatic summarization. It is two step process i.e. the identification of proper nouns and its classification. Identification is concerned with marking the presence of a word/phrase as NE in the given sentences and classification is for denoting role of the identified NE. The NER task was added in Message Understanding Conference (MUC) held in November, 1995 at Los Altos [5][18]. The various approaches of NER are namely- Rule Based, Machine Learning based which includes HMM, Maximum Entropy, Decision Tree, Support Vector Machines and Conditional Random Fields and Hybrid Approach. Although a lot of work has been done in English and other foreign languages like Spanish, Chinese etc with high accuracy but regarding research in Indian languages is at initial stage only. Here a survey of research done till now in English and other foreign and Indian languages are presented. Early systems are making use of handcrafted rule-based algorithms. While modern systems most often use machine learning techniques. Handcrafted rule-based systems usually give good results, however they need months of development by experienced linguists. Whereas machine learning techniques uses a collection of annotated documents to train classifier for the given set of NE classes. According to the specification defined by MUC, the NER tasks generally work on seven types of named entities as listed below: Person Name Location Name Organization Name Abbreviation Time Term Name Measure 2. Previous Work There are several classification methods which are successful to be applied on NER task. Till now, the

261 240 research aiming at automatically identifying named entities in texts forms a vast and heterogeneous pool of strategies, methods and representations. The main approaches to NER are Linguistics approaches and Machine Learning approaches. The Linguistics approach uses rule-based models manually written by linguists. ML based techniques make use of a large amount of annotated training data to acquire high-level language knowledge. Various ML techniques which are used for the NER task are Hidden Markov Model (HMM) [7], Maximum Entropy Model (MaxEnt) [1], Decision Tree [9], Support Vector Machines [20] and Conditional Random Fields (CRFs) [11]. Both the approaches may make use of gazetteer information to build system because it improves the accuracy. Ralph Grishman in 1995 developed a rule-based NER systems which uses some specialized name dictionaries including names of all countries, names of major cities, names of companies, common first names etc[19]. In rule-based approaches, a set of rules or patterns is defined to identify the named entities in a text. Another rule-based NER system is developed in 1996 which make use of several gazetteers like person name, organization name, location names, person names, human titles etc[21]. But the main disadvantages of these rulebased techniques are that these require huge experience and grammatical knowledge of particular languages or domains and these systems are not transferable to other languages. Borthwick in 1999 developed a ML based system i.e. MaxEnt based system[1]. This system used 8 dictionaries. ML based techniques for NER make use of a large amount of NE annotated training data to acquire high level language techniques uses gazetteer lists. A lot of work has been done on NER for English employing the machine learning techniques, using both supervised learning and unsupervised learning. In English language, it is easier to identify NE because of the capitalization of names. Unsupervised learning approaches do not require labeled training data i.e. training requires few seed lists and large unannotated corpora. In unsupervised learning, the goal is to build representations from data. These representations are then be used for data compression, classifying, decision making and other purposes. Unsupervised learning is not a very popular approach for NER and the systems that do use unsupervised learning are usually not completely unsupervised. Collins et. al[6]. discusses an unsupervised model for named entity classification by the use of unlabelled examples of data. Secondly, Koim et. al[10]. Proposes an unsupervised named entity classification models and their ensembles that uses a small-scale named entity dictionary and an unlabeled corpus for classifying named entities. Supervised learning involves using a program that can learn to classify a given set of labeled examples that are made up of the same number of features. The Supervised learning approach requires preparing labeled training data to construct a statistical model. But supervised approaches can achieve good performance only when large amount of high quality training data is available. Supervised approaches are more expensive than unsupervised one, in terms of the time spend to preprocess the training data. Statistical methods such as HMM, Decision Tree Model and Conditional Random Fields have been used. Hidden Markov Model is a generative model. The model assigns the joint probability to paired observation and label sequence. Then the parameters are trained to maximize the joint likelihood of training sets. It is advantageous as its basic theory is elegant and easy to understand. Hence it is easier to implement and analyze. It uses only positive data, so they can be easily scaled. Disadvantage - In order to define joint probability over observation and label sequence HMM needs to enumerate all possible observation sequence. Hence it makes various assumptions about data like Markovian assumption i.e. current label depends only on the previous label. Also it is not practical to represent multiple overlapping features and long term dependencies. Number of parameter to be evaluated is huge. So it needs a large data set for training. Maximum Entropy Markov Models (MEMMs): It is a conditional probabilistic sequence model. It can represent multiple features of a word and can also handle long term dependency. It is based on the principle of maximum entropy which states that the least biased model which considers all know facts is the one which maximizes entropy. Each source state has a exponential model that takes the observation feature as input and output a distribution over possible next state. Output labels are associated with states. Advantages: It solves the problem of multiple feature representation and long term dependency issue faced by HMM. It has generally increased recall and greater precision than HMM. Disadvantages: It has Label Bias Problem. The probability transition leaving any given state must sum to one. So it is biased towards states with lower outgoing transitions. The state with single outgoing state transition will ignore all observations. To handle Label Bias Problem we can change the state-transition Conditional Random Field (CRF): It is a type of discriminative probabilistic model. It has all the advantage of MEMMs without the label bias problem. CRFs are undirected graphical models (also know as random field)

262 241 which is used to calculate the conditional probability of values on assigned output nodes given the values assigned to other assigned input nodes. In Hybrid NER system, approach uses the combination of both rule-based and ML technique and makes new methods using strongest points from each method. It is making use of essential feature from ML approach and uses the rules to make it more efficient. Sirihari et. al. introduce a hybrid system by combination of HMM, MaxEnt, and handcrafted grammatical rules[24]. In the field of NER for English and other European Languages, lots of work has already been done. This is possible because of the main feature in English i.e. the capitalization of names in the text. That is why NER task is achieved with high accuracy. Hybrid Approach: Here several Machine Learning and Rule based systems are combined to improve the accuracy of classifier. Some examples of Hybrid systems are MaxEnt + Rule : Borthwick(1999) 92% f- measure MaxEnt + Rule: Edinburgh Univ % f- measure MaxEnt +HMM + Rule: Srihari et al. (2000) 93.5% f-measure In this field, recent researches are focused on multimedia indexing, unsupervised learning, complex linguistics phenomena and machine translation. Lots of efforts are taken toward semi-supervised and unsupervised approaches to NER motivated by the use of very large collection of texts [8] and the possibility of handling multiple NE types [15]. Complex linguistic phenomena that are common short-coming of current systems are under investigation [17]. The term semi-supervised is relatively recent. The main technique for SSL is called bootstrapping and involves a small degree of supervision, such as a set of seeds, for starting the learning process. Recent experiments in semi-supervised NERC [15] report performance that rival baseline supervised approaches. Features are characteristic attributes of words designed for algorithmic purpose. Following features are most often used for the recognition and classification of named entities. These are defined into three categories i.e. Word-level features List lookup features Document and corpus features Word-level features describe the character makeup of words i.e. the word case, punctuation, numerical value, part-of-speech (POS) and special characters List lookup features can be called also as the term gazetteer, lexicon and dictionary. It include the general list, list of entities such as organization name, first name etc. and the looking into predefined list. Document and corpus features are defined as collection of document content and document structure. Large collection of document (corpora) are also excellent sources of features. These all features together or in different combination helps in generating effective and efficient NER system for different domains or languages. 3. NER for Indian languages NLP research around the world has taken major turn in the last decade with the advent of effective machine learning algorithms and the creation of large annotated corpora for various languages. But not much work has been done in NER for Indian languages because annotated corpora and other lexical resources have started appearing very recently in India. As common feature function like capitalization are not available in Indian languages and due to lack of large labeled dataset and lack standardization and spelling variation, so English NER cannot be directly used for Indian languages. So there arises the need to develop novel and accurate NER system for different Indian languages. 3.1 Characteristic and some problems faced by Hindi and other Indian languages No capitalization Brahmi script- It has high phonetic characteristic which could be utilized by NER system. Non-availability of large gazetteer Lack of standardization and spelling Number of frequently used words (common nouns) which can also be used as names are very large. Also the frequency with which they can be used as common noun as against person name is more or less unpredictable. Lack of labeled data Scarcity of resources and tools Free word order language 3.2 Some points to consider while building NER System Ease to change Portability (domains and language) Scalability Language Resources Cost-effective

263 Performance Evaluation Metrics are: Precision (P): Precision is the fraction of the documents retrieved that are relevant to the user s information need. Precision (P) = correct answers/answers produced Recall (R): Recall is the fraction of the documents that are relevant to the query that are successfully retrieved. Recall (R) = correct answers/total possible correct answers F-Measure: The weighted harmonic mean of precision and recall, the traditional F-measure or balanced F-score is F-Measure = (β 2 +1)PR/( β 2 R+P) β is the weighting between precision and recall typically β=1. When recall and precision are evenly weighted i.e. β=1, F-measure is called F 1 measure. F 1 -Measure = 2PR/(P+R) There is a tradeoff between precision and recall in the performance metric. In IJCNLP-08 workshop on NER for South and South East Asian languages, held in 2008 at IIT Hyderabad, was a major attempt in introducing NER for Indian languages that concentrated on five Indian languages- Hindi, Bengali, Oriya, Telugu and Urdu. The work regarding Telugu language is mentioned in [16]. The evaluation has reported F-Score of 44.91%. The development of a NER system for Bengali language is reported in 2008[2]. Its F-Score is 91.8%. The work of Gali et al, in 2008 reports lexical F- Score of 40.63%, 50.06%, 39.04%, 40.94%, and 43.46% for Bengali, Hindi, Oriya, Telugu, and Urdu respectively [12]. In 2007 discussed the comparative study of Conditional Random Field and Support Vector Machines for recognizing named entities in Hindi language [4]. Indian languages are resource poor languages because of the non-availability of the annotated corpora, name dictionaries, good morphological analyzers etc. That is why high accuracy is not achievable yet. The maximum accuracy for NER in Hindi is reported by Kumar and Bhattacharyya in They achieved an F measure of 79.7% using a Maximum Entropy Markov Model [13]. Among other Indian languages, Punjabi language still lacks behind in this field. A research work is concentrated on NER for Punjabi language. Punjabi is the official language of the Indian state of Punjab. It is also official language of Delhi and ranked 20 th among the language spoken in the world [23]. Among the Indian languages, Punjabi is the one in which the lots of research is going on in this field. Due to the nonavailability of annotated corpora, name dictionaries, good morphological analyzer etc. up to the required measure, Punjabi is the resource poor language like other Indian languages. A recent research on NER for Punjabi language is done using Conditional Random Field (CRF) Approach [25]. It was aimed to develop a standalone system based on CRF approach which can be used with other NLP applications like Machine Translation, Information Retrieval etc. In this paper, 12 named entities are mentioned as in table 1. Table 1: Named Entity Tagset NE Tag Definition NEP(Person) Name of a person NEL(Location) Name of a place, location NEO(Organization) Name of a political organization NED(Designation) NETE(Term) NETP(Title-Person) NETO(Title-Object) NEB(Brand) NEM(Measure) NEN(Number) NETI(Time) NEA(Abbreviation) Name of any designation Name of diseases Name of title coming before the name of person Name of Object Brands Name Any measure Numeric value It include date, month, year etc Name in short form These tagset are used to tag each word in the sentence. Firstly, to find the useful features for NER task and secondly, to find the optimum feature set for the task. The various features which are applied to the NER tasks in this experiment are as follows: Context word feature : Previous and next words of a particular word have been used as a feature. Generally, word window of size 5 or 7 is used. Word suffix and prefix: In this feature, a length of 1 to 4 characters of the current and/or the surrounding words is taken.

264 243 Parts of Speech (POS) Information: A rule-based POS tagger developed at Punjabi University by Gill and Lehal in 2007 is used [14]. It is helpful in tagging the data but with limited accuracy. Some wrong tags are manually corrected for NER task. Named Entity Information: It is the feature in which the NE tag of the previous word is considered. It is the dynamic feature. Gazetteer Lists: Due to the scarcity of resources in electronic format for Punjabi language, so the gazetteer lists are prepared manually from websites and newspaper available online. Seven different lists are prepared such as: Person-Prefix First-Name Middle-Name Last-Name Location-Name Month Name Day Name The F-score is calculated for the different use of features to obtain the optimal feature set. An overall F-score of 80.92% achieved for the Punjabi NER. The F-score has different value for the different NE tags. This means NER systems can be changed according to the type of NE tags required. The performance can be improved by improving gazetteer lists. 4. Conclusions The Named Entity Recognition field has been thriving for more than fifteen years. It aims at extracting and classifying mentions of rigid designators, from text, such as proper names and temporal expressions. In this survey, we have shown the previous work done in English and other European languages. A survey is given on the work done in Indian Languages i.e. Telugu, Hindi, Bengali, Oriya and Urdu. An overview of the techniques employed to develop NER systems, documenting the recent trend away from hand-crafted rules towards machine learning approaches. Handcrafted systems provide good performance at a relatively high system engineering cost. When supervised learning is used, a prerequisite is the availability of a large collection of annotated data. Such collection are available from the evaluation forums but remain rather rare and limited in domain and language coverage. Recent studies in the field have explored semisupervised and unsupervised learning techniques that promise fast deployment for many entities types without the prerequisite of an annotated corpus. Here also provided an overview of the evaluation methods that are in the use of NER accuracy. We have listed and categorized the features that are used in recognition of NE. The use of an expressive and varied set of features turns out to be just as important as the choice of machine learning algorithms. And finally the survey on the NER for Punjabi language is given. In it the working of an approach is explained. 5. Future work The performance can further be improved by improving gazetteer lists. Analyzing the performance using other methods like Maximum Entropy and Support Vector Machines Comparing the results obtained by using different approaches and calculating the most accurate approach for it. Improve the performance of each NE tag to make it overall more accurate. References [1] Andrew Borthwick Maximum Entropy Approach to Named Entity Recognition Ph.D. thesis, New York University. [2] Asif Ekbal, Sivaji Bandyopadhyay. Bengali Named Entity Recognition using Support Vector Machine in the proceedings of the IJCNLP-08 workshop on NER for South and South East Asian Languages, pages 51-58, Hyderabad, India. [3] Asif Ekbal, Rejwanul Haque, Amitava Das, Venkateshwar Poka and Sivaji Bandyopadhyay , Language Independent Named Entity Recognition in Indian Languages in the proceedings of the IJCNLP- 08 Workshop on NER for South and South East Asian Languages, pages 33-40, Hyderabad, India. [4] Awaghad Ashish Krishnarao, Himanshu Gahlot, Amit Srinet, D.S.Kushwaha, A Comparison of Performance of Sequential Learning Algorithms on the task of Named Entity Recognition for Indian Languages in the proceedings of 9 th International Conference on computer Science. Pages Baton Rouge, LA, USA. [5] Charles L. Wayne , A snapshot of two DARPA speech and Natural Language Programs in the proceedings of workshop on Speech and Natural Languages, pages , Pacific Grove, California. Association for Computational Linguistics. [6] Collins, Michael and Y. Singer Unsupervised models for Named Entity Classification, in the proceedings of the Joint SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora.

265 244 [7] Daniel M. Bikel, Scott Miller, Richard Schwartz and Ralph Weischedel Nymble: a highperformance learning name-finder in the proceedings of the fifth conference on Applied natural language processing, pages , San Francisco, CA, USA Morgan Kaufmann Publishers Inc. [8] Etzioni, Oren; Cafarella, M; Downey, D.; Popescu, A.- M.; Shaked, T.; Soderland, S.; Weld, D.S.; Yates, A Unsupervised Named-Entity Extraction from the Web: An Experimental Study in Artificial Intelligence 165. Pages , Essex: Elsevier Science Publishers. [9] Hideki Isozaki Japanese named entity recognition based on a simple rule generator and decision tree learning in the proceedings of the Association for Computational Linguistics, pages India. [10] J. Kim, I. Kang, K. Choi, Unsupervised Named Entity Classification Models and their Ensembles, in the proceedings of the 19 th International Conference on Computational Linguistics, [11] John D. Lafferty, Andrew McCallum, and Fernando C. N. Pereira Conditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data in the proceedings of International Conference on Machine Learning, pages , Williams College, Williamstown, MA, USA. [12] Karthik Gali, Harshit Surana, Ashwini Vaidya, Praneeth Shishtla and Dipti Misra Sharma , Aggregating Machine Learning and Rule Based Heuristics for Named Entity Recognition in the proceedings of the IJCNLP-08 Workshop on NER for South and South East Asian Languages, pages 25-32, Hyderabad, India. [13] Kumar N. and Bhattacharyya Pushpak Named Entity Recognition in Hindi using MEMM in the proceedings of Technical Report, IIT Bombay, India. [14] Mandeep Singh Gill, Gurpreet Singh Lehal and Shiv Sharma Joshi, Parts-of-Speech Tagging for Grammar Checking of Punjabi in the Linguistics Journal Volume 4 Issue 1, pages [15] Nadeau, David; Turney, P.; Matwin, S Unsupervised Named Entity Recognition; Generating Gazetteers and Resolving Ambiguity in the proceedings of Canadian Conference on Artificial Intelligence. [16] Praneeth M Shishtla, Karthik Gali, Prasad Pingali and Vasudeva Varma Experiments in Telugu NER: A conditional Random Field Approach in the proceedings of the IJCNLP-08 Workshop on NER for South and South East Asian Languages, pages , Hyderabad, India. [17] Poibeau, Thierry Dealing with Metonymic Readings of Named Entities in the proceedings of Annual Conference of the Cognitive science Society. [18] R. Grishman, Beth Sundheim Message Understanding Conference-6: A Brief History in the proceedings of the 16 th International Conference on Computational Linguistics (COLING), pages , Center for Sprogteknologi, Copenhagen, Denmark. [19] R. Grishman The NYU system for MUC-6 or Where s the Syntax in the proceedings of Sixth Message Understanding Conference (MUC-6), pages , Fairfax, Virginia. [20] Takeuchi K. and Collier N Use of Support Vector Machines in extended named entity recognition in the proceedings of the sixth Conference on Natural Language Learning (CoNLL-2002), Taipei, Taiwan, China. [21] Wakao T., Gaizauskas R. and Wilks Y Evaluation of an algorithm for the Recognition and Classification of Proper Names, in the proceedings of COLING-96. [22] [23] [24] Srihari R., Niu C. and Li W A Hybrid Approach for Named Entity and Sub-Type Tagging in the proceedings of the sixth Conference on Applied Natural Language Processing. [25] Amandeep Kaur, Gurpreet S. Josan and Jagroop Kaur. Named Entity Recognition for Punjabi: A Conditional Random Field Approach in the proceedings of 7 th International Conference on Natural Language Processing, Macmillan Publishers, India. First Author Darvinder Kaur is Assistant Professor in Computer Science and Engineering Department at Lovely Professional University, Phagwara. She has done M.E. in Computer Science and Engineering from University Institute of Engineering and Technology, Panjab University, Chandigarh in She has done B.Tech in Computer Science and Engineering from Guru Nanak Dev Engineering College, Ludhiana in 2008.

266 245 Vishal Gupta is Assistant Professor in Computer Science & Engineering Department at University Institute of Engineering & Technology, Panjab university Chandigarh. He has done MTech. In computer science & engineering from Punjabi University Patiala in He was among university toppers. He secured 82% Marks in MTech. Vishal did his BTech. in CSE from Govt. Engineering College Ferozepur in He is also pursuing his PhD in Computer Sc & Engg. Vishal is devoting his research work in field of Natural Language processing. He has developed a number of research projects in field of NLP including synonyms detection, automatic question answering and text summarization etc. One of his research paper on Punjabi language text processing was awarded as best research paper by Dr. V. Raja Raman at an International Conference at Panipat. He is also a merit holder in 10th and 12th classes of Punjab School education board. in professional societies. The photograph is placed at the top left of the biography.

267 246 A Framework for Prefetching Relevant Web Pages using Predictive Prefetching Engine (PPE) Jyoti 1, A K Sharma 2 and Amit Goel 3 1 Dept of Computer Science, YMCA University of Sc. and Tech., Faridabad, Haryana, India 2 Dept of Computer Science, YMCA University of Sc. and Tech., Faridabad, Haryana, India 3 Manager, Evalueserve, Gurgaon, Haryana, India Abstract This paper presents a framework for increasing the relevancy of the web pages retrieved by the search engine. The approach introduces a Predictive Prefetching Engine (PPE) which makes use of various data mining algorithms on the log maintained by the search engine. The underlying premise of the approach is that in the case of cluster accesses, the next pages requested by users of the Web server are typically based on the current and previous pages requested. Based on same, rules are drawn which then lead the path for prefetching the desired pages. To carry out the desired task of prefetching the more relevant pages, agents have been introduced. Keywords: Predictive-Prefetching. 1. Introduction It is indisputable that recent explosion of World Wide Web has transformed not only the discipline of computerrelated sciences but also the lifestyles of people and the economies of the countries. Web server is the single most piece of software that enables any kind of web activity. Since its inception, web server has always taken the form of a daemon process. It takes http request, interprets it and serves the file back. As web services are increasingly becoming popular, network congestion and server overloading have become significant problems. To overcome these problems, efforts are being made continuously to increase the web performance. Web caching is recognized as one of the effective techniques to alleviate the server bottleneck and reduce network traffic, thereby reducing network latency. The basic idea is to cache requested pages at the server so that they don t have to be fetched again. Although web cache schemes reduce the network and I/O bandwidth consumption, they still suffer from a low hit rate, stale data and inefficient resource management. [1] shows that an inefficient web cache management caused a major news web site crash, also called the Slashdot effect. Web prefetch schemes overcome the limitation of web cache mechanisms through pre-processing contents before a user request comes. Web prefetch schemes expect future requests through web log file analysis and prepare the expected requests before receiving it. Compared with web cache schemes, web prefetch schemes focus on the spatial locality of objects when current requests are related with previous requests. Web prefetch schemes increase the bandwidth utilization and reduce or hide the latency due to bottleneck at web server. But prefetching scheme should be carefully chosen as a wrong prefetching system can cause major network bandwidth bottlenecks rather than reducing the web-user-perceived latency. The organization of the paper is as follows. Section 2 discusses the related work. Section 3 introduces the proposed framework with the required references to the various components of the same. Subsection 3.1 talks about the components of the proposed work followed by the flow process of PPE in subsection 3.2 while subsection 3.3 illustrates the whole process with the help of the flowchart. Section IV concludes the paper followed by the references.

268 Related work Web users can experience response times in the order of several seconds. Such response times are often unacceptable, causing some users to request the delayed documents again. This, in turn, aggravates the situation and further increases the load and the perceived latency. Caching is considered an effective approach for reducing the response time by storing copies of popular Web documents in a local cache, a proxy server cache close to the end user, or even within the Internet. However, the benefit of caching diminishes as Web documents become more dynamic [2]. A cached document may be stale at the time of its request, given that most Web caching systems in use today are passive (i.e., documents are fetched or validated only when requested). Prefetching (or proactive caching) aims at overcoming the limitations of passive caching by proactively fetching documents in anticipation of subsequent demand requests. Several studies have demonstrated the effectiveness of prefetching in addressing the limitations of passive caching (e.g., [3, 4, 5, 6, 7, 8, 9, 10, 11, 12, and 13]). Prefetched documents may include hyperlinked documents that have not been requested yet as well as dynamic objects [14, 11]. Stale cached documents may also be updated through prefetching. In principle, a prefetching scheme requires predicting the documents that are most likely to be accessed in the near future and determining how many documents to prefetch. Most research on Web prefetching focused on the prediction aspect. In many of these studies (e.g., [15, 10]), a fixed-threshold-based approach is used, whereby a set of candidate files and their access probabilities are first determined. Among these candidate files, those whose access probabilities exceed a certain prefetching threshold are prefetched. Other prefetching schemes involve prefetching a fixed number of popular documents [9]. Teng et. al [16] proposed the Integration of Web Caching and Prefetching (IWCP) cache replacement policy, which considers both demand requests and prefetched documents for caching based on a normalized profit function. The work in [17] focuses on prefetching pages of query results of search engines. In [18], the authors proposed three prefetching algorithms to be implemented at the proxy server: (1) the hit-rate-greedy algorithm, which greedily prefetches files so as to optimize the hit rate; (2) the bandwidth-greedy algorithm, which optimizes bandwidth consumption; and (3) the H/B-greedy algorithm, which optimizes the ratio between the hit rate and bandwidth consumption. The negative impact of prefetching on the average access time was not considered. Most of the above works rely on prediction algorithms that compute the likelihood of accessing a given file. Such computation can be done by employing Markovian models [19, 10, 20, and 21]. Other works rely on data mining for prediction of popular documents [22] [23] [24] [25]. Numerous tools and products that support Web prefetching have been developed [26], [27, 28, and 29]. Wcol [30] prefetches embedded hyperlinks and images, with a configurable maximum number of prefetched objects. PeakJet2000 [29] is similar to Wcol with the difference that it prefetches objects only if the client has accessed the object before. NetAccelerator [28] works as PeakJet2000, but does not use a separate cache for prefetching as in PeakJet2000. Google s Web accelerator [31] collects user statistics, and based on these statistics it decides on what links to prefetch. It also can take a prefetching action based on the user s mouse movements. Web browsers based on Mozilla Version 1.2 and higher also support link prefetching [32]. These include Firefox [26], FasterFox [33], and Netscape [27]. In these browsers, Web developers need to include html link tags or html meta-tags that give hints on what to prefetch. In terms of protocol support for prefetching, Davison et al. [34] proposed a prefetching scheme that uses a connectionless protocol. They assumed that prefetched data are carried by low-priority datagrams that are treated differently at intermediate routers. Although such prioritization is possible in both IPv6 and IPv4, it is not yet widely deployed. Kokku et al. [35] proposed the use of the TCP-Nice congestion control protocol [36] for low-priority transfers to reduce network interference. They used an end-to-end monitor to measure the server s spare capacity. The reported results show that careful prefetching is beneficial, but the scheme seems to be conservative because it uses an additive increase (increase by 1), multiplicative decrease policy to decide on the amount of data to prefetch. Crovella et. al [37] showed that a rate-control strategy for prefetching can help reduce traffic burstiness and queuing delays. Most previous prefetching designs relied on a static approach for determining the documents to prefetch. More specifically, such designs do not consider the state of the network (e.g., traffic load) in deciding how many documents to prefetch. For example, in threshold-based schemes, all documents whose access probabilities are greater than the prefetching threshold are prefetched. As shown in this paper, such a strategy may actually increase the average latency of a document. 3. Proposed work The general architecture of a common Web search engine contains a front-end process and a back-end process. In the front-end process, the user enters the search keywords into the search engine interface, which is usually a Web page with an input box. The application then parses the search request into a form that the search engine can understand, and then the search engine executes the search operation on the index files. After ranking, the search

269 248 engine interface returns the search results to the user. In the back-end process, a spider or robot fetches the Web pages from the Internet, and then the indexing subsystem parses the Web pages and stores them into the index files. The search engine retrieves the web pages according to the user query. Since relevancy is a subjective term, the search results may have varying degree of relevancy for different set of users. Given this fact, there is an opportunity to significantly improve the relevancy of search results for a well defined set of users (example, employees of the same organisation), whose search habbits are largely homogenous. The proposed work introduces the Predictive Prefetching Engine (PPE) which sits behind the search engine interface. The intent of introducing the PPE [38] is that it will increase the relevancy of the pages returned by the search engine according to the demand of the particular set of users which are termed as group clients. PPE also prefetches the pages if it lies in the rule-database that is generated by applying the various data mining operations on the group-client-log. This log is maintained by the search engine on the request of the various organisations which are assigned a particular set of IP addesses by the Internet Service Providers. The interaction of the PPE with the user and the process of retrieving the relevant web pages from the WWW is explained in the next subsection. 3.1 Components of the Proposed Work 1. Search Engine Interface: It is the part of the search engine s front end and is basically a web page with the input box. The user enters its query conataining the keywords into this input box and hits the search button. 2. IP Matcher: It extracts the IP address from the query coming from a particular user. This IP address is then matched with the particular range of IP addresses for which different Group-Client-Agents (GCAs) are defined. Once the GCA is identified, it gets activated. 3. Group-Client-Agent (GCA): As the name suggests, it is an agent. GCA plays the crucial role as it will work on PPE. There are be n GCA s for n group-clients and hence each GCA will have a corresponding PPE to work upon. One group-client refers to a group of users within one organisation. Every organisation is assigned a unique set of IP addresses. These IP addresses will form a part of one group-client. 4. Group-Client-Log(GC-Log): This log is maintained by the search engine on the Group-client s request. The format of the log is same as that of the web server maintained by the search engine and contains every entry from that particular group-client. Each record in the log file contains the client s IP address, the date and time the request is received, the requested object and some additional information such as protocol of request, size of the object etc. Fig. 1 Framework for retrieving the relevant web pages from WWW using PPE 5. Clean Log: This log is cleaned by removing all the image files like.jpg and.gif from GC-Log as they yield no productive information about the path followed by the user in a particular session. 6. RST Clusters: The clean log is then treated to find the user sessions. A session is the sequence of pages viewed and actions taken by a single user during a defined period of time i.e. 30 minutes. Analyzing the web access log and user sessions, user behavior can be understood. These sessions are then operated upon by the clustering technique known as Rough Set Clustering. The purpose of clustering the sessions is to reduce the search space for applying the various datamining operations. RST operates on the principle of indiscernibility which is defined as equivalence between the objects. RST is chosen as the clustering technique as it aids in decision making in the presence of uncertainty. The result of applying RST is the lower approximation set which contains all the user sessions which definitely contain the target set [39]. 7. Rule Generator: By making use of rough set clustering, those user sessions were deduced from the web log in which the user spends his quality time. These sessions are in the Lower Approximation set [40]. These sessions are then fed to Rule generator phase of PPE where k-order markov predictors are applied onto these user sessions. It is important to formulate the value of k so that its value is decided dynamically as keeping its value low or high have

270 249 their own drawbacks. So, the optimum value of k has to be chosen. Here, the minimum threshold that will be used in deciding rules would be half of the maximum number of time a particular sequence of web pages is used. i.e. if the maximum time a particular page sequence called is 6 then minimum threshold to consider other page sequences must be 6/2=3 and k is this minimum threshold. Thus, k is being decided dynamically. The output of this phase is the rules of the form Di > Dj. 8. Rule Repository: The rules formed in the last phase are then stored in the repository. 9. Database: This database contains the URLs of all the pages whose references are stored in the rule repository. The database is enriched by the URLs of rules from all the n PPE s by the GCA s Page loader: Its job is to prefetch the pages populated in the hint-list by the GCA onto the client s cache. forwarded by the GCA to the crawler. The crawler then crawl the web pages from the WWW and after indexing, add them to the Database of URLs. 8. Once the GCA has populated its hint-list with the web pages, it sends the signal to the page loader. The page loader then prefetches the client s cache with the respective GCA s hint list Flowchart of the Proposed Work This subsection discusses the whole process of how the user is returned with the prefetched pages. 3.2 FlowProcess for Prefetching the desired documents Once the IP Matcher identifies the GCA according to the client machine, GCA gets activated and starts working on the prefetching scheme which is as follows: 1. Let the request be for document A. 2. The agent scans the rule database* for the rules of the form A X for some document X. 3. The agent then scans the database for every rule or part of the rule which has X in its sequence (e.g. A Y X Z). The only exception to this scan would be in the case of X being the last document in the sequence. 4. As it scans, the agent brings the URLs of all the documents that succeed X from the Database of URLs to its hint list and accordingly prefetch them to the client s cache. 5. The agent continues the scan and populates the hint list till such time the user requests for a web page which doesn t appear in the sequence. In such case, the agent cleans up the hint list and starts afresh. (Step 2). 6. If the GCA finds two rules with the same head but each having a different tail, then the GCA applies the subsequence association rules to find their confidence. The confidence is calculated based on their past history. The rule whose past history generates the maximum confidence is considered by GCA for prefetching. This helps in saving the network bandwidth which is generally considered an issue in the design of the prefetching mechanism. 7. If in case the document A doesn t match as the head of the rule in the Rule-Repository, the request is Fig. 2: Flowchart for Prefetching the Web pages according to the rules formed by PPE Fig. 2 is the abstracted version of the whole process. It is the flowchart that shows the path that is followed right from when the user enters the query in the search engine interface, how a particular GCA (Group-Client-Agent) gets activated and the tasks it then performs to serve its client with the prefetched pages. 4. Transaction Processing Phases of PPE The overall processing of the transactions from the calculation of user accesses to the generation of rules to the Prefetching of the pages into the cache is occurring in three main phases as shown in Fig 3. The step wise working of these phases is as follows: * The rule database can be organized using some indexing scheme.

271 250 Fig. 3 Transaction Processing Phases The step wise working of these phases is as follows: 1. Clustering User Sessions: In this phase the user sessions are clustered. To perform this task, two subtasks need to be performed. They are the identification of the user transactions from the GC- Log and then applying RST (Rough Set clustering) over the user sessions to cluster those sessions which definitely contain the target set. a. Identification of the user transactions: The foremost thing for the determination of the user transactions is the identification of the user sessions from the log file. The objective of user session is to separate independent accesses made by different users or by the same user at distant points in time [41, 42]. b. RST Clusters: A rough set, first described by Zdzisław I. Pawlak, is a formal approximation of a crisp set (i.e., conventional set) in terms of a pair of sets which give the lower and the upper approximation of the original set. Formally, an information system is a pair A = (U, A) where U is a non-empty, finite set of objects called the universe and A is a non-empty, finite set of attributes on U.With every attribute a A, a set V a is associated such that a: U V a. The set V a is called the domain or value set of attribute a. Indiscernibility is core concept of RST and is defined as equivalence between objects. Objects in the information system about which we have the same knowledge form an equivalence relation. The equivalence relation has the following properties. If a binary relation R X * X which is reflexive (i.e. an object is in relation with itself xrx), symmetric (if xry then yrx) and transitive (if xry and yrz then xrz) is called an equivalence relation.) Formally any set B A, there is associated an equivalence relation called B-Indiscernibility relation defined as follows: IND A (B) = {(x, x ) U 2 a B a(x) = a (x )} If (x, x ) IND A (B), then objects x and x are indiscernible from each other by attributes from B. Equivalence relations lead to the universe being divided into equivalence class partition and union of these sets make the universal set. Target set is generally supposed by the user. Lower approximation is the union of all the equivalence classes which are contained by the target set. The lower approximation is the complete set of objects that can be positively (i.e., unambiguously) classified as belonging to target set X. The P-upper approximation is the union of all equivalence classes which have non-empty intersection with the target set. It represents the negative region, containing the set of objects that can be definitely ruled out as members of the target set. 2. Rule Determiner: Once the user sessions are clustered as lower approximation set, the next step is to determine the rules. These rules will let know which pages are to be prefetched. To determine the rules, markov predictors will be used. E.g. if S= p 1,, p n is a sequence of accesses (called a transaction) made by a user, then the conditional probability that the next access will be p n+1 is P(p n+1 p 1,, p n ). Therefore, given a set of transactions, rules of the form: p 1,, p n = p n+1 (1) can be derived, where P(p n+1 p 1,, p n ) is equal to or larger than the user defined cut-off threshold value T c. The left part of the rule is called the head and the right part is called the body. The body of the rule can also be any length larger than one. E.g. rules of the form p 1,, p n = p n+1,, p n+m (2) In this case, P(p n+1,, p n+m p 1,, p n), has to be larger than T c. The dependency of the forthcoming accesses on past accesses defines a Markov Chain. The number of past accesses considered in each rule for the calculation of the corresponding conditional probability is called the order of the rule. E.g. the order of the rule A, B=>C is 2. The predictive web prefetching algorithm can be defined as a collection of 1, 2 k-order Markov Predictors. An k-order Markov predictor is defined to be a scheme for the calculation of conditional probabilities P(p n+1,, p n+m p 1,, p n ) between document accesses and the determination of the rules of the form (2). The head of the each rule has a size

272 251 equal to n and the body of each rule has the size equal to m. The job of determining the rules is performed by the rule generator component which are then stored in the Rule repository component of the proposed framework as shown in Fig Rule Activator: After the determination of the rules of the form (2), the next requirement is for the activation mechanism. The rule activator phase accomplishes the task of finding the prefetched pages from the corresponding rules. This phase makes use of the GCA (group-client agent) which matches the user s request for the documents with the heads of the rules. If the suitable match is found, it will prefetch the documents found in the tail of the corresponding rule. 5. Empirical Results Since relevancy is a very subjective term, to plot it in the form of a graph is difficult. But still keeping some relevancy factor as a baseline, the experimental setup tries to prove the point that For any given set of keywords, no matter how many times they are entered in the search engine within a given time frame, it shows the same position of the URLs that it fetches from the WWW while prefetching if done carefully can greatly help in repositioning the desired URL. For user, to reach from one relevant page to other relevant page, he has to go through all the series of URLs in between. But since the proposed mechanism works over the Rule Repository formed from the past history of the users access patterns, the rules directly help in prefetching the desired page in the client s cache. Thus, he need not traverse all the in between URLs to reach the desired page. In the following example, for a particular keyword set mobile agents based information retrieval for web mining entered, say the search engine returns 50 URLs in total as noted in table 1. It is known that general search engine returns the most relevant pages on first page with first URL having the highest relevancy and so on. Assuming there are 10 URLs on each page so in total search engine returns 5 pages. The first URL will have highest relevancy with the relevancy factor of 50 and the 50 th URL will have the lowest relevancy factor of 1. Ke yw or d URL No. Relevancy factor Table 1 URL K1 U vwbd9bxmvcrym.pdf U2 49 portal.acm.org/citation.cfm?id= U Mining-09.ppt U ent.showcfp?eventid=11144 U5 46 scialert.net/fulltext/?doi=jai &org=11 U sue2-1/kosala.pdf U7 44. U50 1 iaup.org/.../7-summary-reportof-the-international-conferenceon-ict-and-knowledgeengineering-2009 Table 2 shows the rule set obtained by PPE. Table 2 S. No Rules 1 U3->U37 2 U2->U12->U18 3 U1->U23->U33 4 U17->U21 Following subsection compares the two approaches: 1. When search engine doesn t perform prefetching. 2. When search engine performs prefetching. Graph 1

273 252 Thus, U37 which was placed at position 7 on page 3 (position 37 in total listing) by general search engine has moved to position 4 on page 1 when search engine employs prefetching. The following charts 1 & 2 illustrate how the search space for the user reduces drastically if the search engine employs prefetching. 50 U R L 37 C L I C K Page number Chart 1: Search space requirement of general search engine Graph 1 plots the various URLs fetched by the search engine against their relative relevancy when search engines does not employ prefetching. So U1 (URL1) being the most relevant URL according to search engine drags the first position on page 1. Each page contains 10 URLs. So U11 is the first URL on page 2 and U37 is the 7 th URL on page 3. Thus, with the increasing page numbers, the relevancy of the URLs decreases according to the normal working of the search engine. The proposed work on the other hand determines the rules which determine the next likely page to be accessed by the user. It may be observed from the first rule of Table 2, which says that if user accesses U3 then the next URL likely to be accessed by the user is U37. In graph 1(a), the orange bar shows the U3 (the first page accessed by the user according to the keywords entered). For him to access U37, he has to traverse through all the URLs on page 1, page 2 and page 3. Also, according to search engine, the relevancy factor associated with U37 is 14 shown with green bar. Graph 1(b) shows how the relevancy of U37 drastically improves if search engine employs prefetching. The comparison can be seen clearly in graph 1(c) with the relevant orange and green bars coming adjacent to each other. Search area to be covered by user as per results of general search engine is SA= 37*4= 148 Search area to be covered by user if search engine employs prefetching is SA = 4*1=4 SA /SA= 4/148= 1/37 If we calculate the ratio of search spaces of both the search engines, it can be seen that search space reduces drastically i.e. by 1/37 th, if search engine employs prefetching which proves the point. U R L C L I C K th URL Page number

274 253 Chart 2: Search space requirement if search engine employs prefetching CONCLUSION The search engine retrieves the web pages for the general user. Since relevancy is a subjective term, the search results may have varying degree of relevancy for different set of users. The proposed work introduces the PPE for retrieving the web pages for the particular set of users named group-clients whose surfing pattern is logged in the CG-log maintained by the search engine only. Since, these group-clients reflect a particular behaviour over a period of time, PPE encaches the same to return not only the relevant web pages but also prefetches them according to their history. Thus, PPE while prefetching the web pages makes sure that the network bandwidth is not wasted. REFERENCES [1] I. Ari, B. Hong, E. L. Miller, S. A. Brandt, and D. Long. Managing flash crowds on the internet, In Proc. Of MOSCOTS, [2] F. Douglis, A. Feldmann, and B. Krishnamurthy. Rate of change and other metrics: a live study of the world wide web. In Proceedings of USENIX Symposium on Internet Technologies and Systems, [3] A. Bestavros. Using speculation to reduce server load and service time on the WWW. In Proceedings of the 4th ACM International Conference on Information and Knowledge Management, pages , [4] M. Crovella and P. Barford. The network effects of prefetching. In Proceedings of IEEE INFOCOM Conference, pages , [5] D. Duchamp. Prefetching hyperlinks. In Proceedings of 2nd USENIX Symposium on Internet Technologies and Systems, pages , [6] L. Fan, P. Cao, W. Lin, and Q. Jacobson, Web prefetching between low-bandwidth clients and proxies: Potential and performance, In Proceedings of ACM SIGMETRICS Conference on Measurment and Modeling of Computer Systems, pages , May [7] T. Kroeger, D. Long, and J. Mogul. Exploring the bounds of web latency reduction from caching and prefetching. In USENIX Symposium on Internet Technologies and Systems, 1997 [8] T. S. Loon and V. Bharghavan. Alleviating the latency and bandwidth problems in WWW browsing. In Proceedings of USENIX Symposium on Internet Technologies and Systems, pages , [9] E. Markatos and C. Chronaki. A top-10 approach to prefetching on the web. In Proceedings of the INET Conference, [10] V. Padmanabhan and J. Mogul. Using predictive prefetching to improve world wide web latency. In Proceedings of the ACM SIGCOMM Conference, pages 26 36, [11] S. Schechtera, M. Krishnanb, and M. Smithc. Using path profiles to predict http requests. In Proceedings of the 7th Internationa World Wide Web Conference, pages , April [12] A. Venkataramani, P. Yalagandula, R. Kokku, S. Sharif, and M. Dahlin. The potential costs and benefits of long-term prefetching for content distribution. In The Sixth Web Caching and Content Distribution Workshop, 2001 [13] I. Zukerman, D. Albrecht, and A. Nicholson. Predicting users requests on the WWW. In Proceedings of the 7th International Conference on User Modeling, pages , 1999 [14] A. B. Pandey, J. Srivastava, and S. Shekhar. Web proxy server with intelligent prefetcher for dynamic pages using association rules. Technical , University of Minnesota, Computer Science and Engineering, January [15] A. Bestavros. Using speculation to reduce server load and service time on the WWW. In Proceedings of the 4th ACM International Conference on Information and Knowledge Management, pages , [16] W.-G. Teng, C.-Y. Chang, and M.-S. Chen. Integrating web caching and web prefetching in client-side proxies. IEEE Transactions on Parallel and Distributed Systems, 16(5): , May [17] R. Lempel and S. Moran. Optimizing result prefetching in web search engines with segmented indices. ACM transactions on Internet Technology, 4(1):31 59, February [18] B. Wu and A. D. Kshemkalyani. Objective-optimal algorithms for long-term Web prefetching. IEEE Transactions on Computers, 55(1):2 17, [19] M. Deshpande and G. Karypis. Selective Markov models for predicting Web page accesses. ACM Transactions on Internet Technology, 4(2): , May [20] R. Sarukkai and S. Clara. Link prediction and path analysis using Markov chains. In Proceedings of the Ninth international World Wide Web Conference, pages , Amsterdam, The Netherlands, [21] T. Palpanas and A. Mendelzon. Web prefetching using partial match prediction. In Proceedings of Web Caching Workshop, San Diego, California, March [22] J. Pitkow and P. Pirolli. Mining longest repeated subsequences to predict world wide web surfing. In Proceedings of the Second USENIX Symposium on Internet Technologies and Systems, Boulder, Colorado, USA, October [23] Q. Yang, H. H. Zhang, and T. Li. Mining web logs for prediction models in WWW caching and prefetching. [24] B. Lan, S. Bressan, B. C. Ooi, and K.-L. Tan. Rule-assisted prefetching in web-server caching. In Proceedings of the ninth international conference on Information and knowledge management, pages , McLean, Virginia, United States, [25] A. Nanopoulos, D. Katsaros, and Y. Manolopoulos. Effective prediction of web-user accesses: a data mining approach. In Proceedings of web usage analysis and user profiling workshop, pages , San Fransisco, CA, [26] [27] [28] Web 3000 Inc. (NetSonic InternetAccelerator).

275 254 [29] PeakSoft corporation. PeakJet 2000 web page [30] [31] [32] prefetching faq/. [33] [34] B. Davison and V. Liberatore. Pushing politely: Improving web responsiveness one packet at a time. Performance Evaluation Review,28(2):43 49, September [35] R. Kokku, P. Yalagandula, A. Venkataramani, and M. Dahlin. NPS: A non-interfering deployable web prefetching system. In USENIX Symposium on Internet Technologies and Systems, [36] A. Venkataramani, R. Kokku, and M. Dahlin. TCP Nice: A mechanism for background transfers. In Proceedings of the 5th Symposium on Operating Systems Design and implementation, volume 36, pages , [37] M. Crovella and P. Barford. The network effects of prefetching. In Proceedings of IEEE INFOCOM Conference, pages , [38] Jyoti, A K Sharma, Amit Goel A Framework for Extracting Relevant Web Pages from WWW using web mining, In Proc of International journal of Computer Society and Network security, 2007 Seoul, Korea. [39] Jyoti, A K Sharma, Amit Goel, Payal Gulati A novel approach for finding sessions using RST, ACT 2009, ieee conference proceedings, Kerela, India. [40] Jyoti, A K Sharma, Amit Goel A Novel Approach to Determine the Rules for Web Page Prediction using Dynamically Chosen K-Order Markov Models, In Proc of IEEE sponsored International Conference on Advances and Emerging trends in Computing Technologies, 2010 [41] M.S. Chen, J.S. Park, and P.S. Yu, Efficient Data Mining for Path Traversal Patterns, IEEE Trans. Knowledge and Eng., Vol.10, n0.2, pp , Apr [42] R. Cooley, B. Mobasher and J.Srivastva, Data Preparation for mining World Wide Web Browsing Patterns, Knowledge and information Systems(KAIS), Vol. 1, no.1, pp. 5-32, Feb [43] Jyoti, A K Sharma, Amit Goel A novel approach for Extracting Relevant Web pages from WWW using Data Mining, Proc. Common Ground Publishers, Australia, 2006 First Author: Ms Jyoti received her M.Tech with Hons. from Maharishi Dayanand University, Faridabad (India) in the year She is currently pursuing her Ph.D in Internet Technology. Since July 2004, she is serving as Asst. Prof. in Computer Engg. at YMCA University of Science and Technology, Faridabad (India). She has published more than 10 papers in various National/International Journals / Conferences. She has guided more than 20 B.Tech/ M.tech Dissertations. Her area of interests is Data Mining, Operating Systems and Object Oriented Technologies. Second Author Prof. A. K. Sharma received his M.Tech. (Computer Sci. & Tech) with Hons. From University of Roorkee in the year 1989 and Ph.D (Fuzzy Expert Systems) from JMI, New Delhi in the year From July 1992 to April 2002, he served as Assistant Professor and became Professor in Computer Engg. at YMCA Institute of Engineering Faridabad in April He obtained his second Ph.D. in IT from IIIT & M, Gwalior in the year His research interests include Fuzzy Systems, Object Oriented Programming, Knowledge Representation and Internet Technologies. Third Author:Dr. Amit Goel received his M.E. with Hons. from National Institute of Technology (NIT), Trichy (India) in the year 2003 and Ph.D.(Comp. Engg.) from M.D.U Rohtak in the year From July 03 to Nov. 06, he has served as Lecturer in Computer Engg. at YMCA Institute of Engineering, Faridabad (India). Since Nov 06, he is working as Patent Analyst at Intellevate (india) private limited. The nature of job pertains to Prior Art Search, Invalidity Search, Claim Mapping etc. He did his Ph.D. in the area of Mobile Ad hoc Networks and designed a secured multicast based power efficient routing protocol. He has published more than 20 research papers in International / National Journals and Conferences. He has guided more than 30 M.Tech / B.Tech dissertations. Presently, he is guiding a Ph.D. student in the area of Internet Technology. His research interest includes Computer Networks, Mobile & Wireless Communication, Database Management Systems, Algorithms and Internet Technology...

276 255

277 255 Genetic Algorithms as Virtual User Managers Rasmiprava Singh 1, Snehalata Baede 2 and Sujata Khobragade3 1 National Institute of Technology,Raipur CG,India 2 National Institute of Technology,Raipur CG,India 3 National Institute of Technology,Raipur CG,India Abstract A growing issue in genetic algorithm research involves understanding the paths through the solution space that are explored. This work presents a software design architecture that can aid in the explanation process. This architecture encourages more complete user interfaces on the problem domain application, which facilitates the integration of a genetic algorithm, at the same time, taking advantage of substantial code re-use. The steps taken to increase the usability of the problem domain application can aid in the visualization of the evolutionary paths explored by the genetic algorithm. This effort on the designer's part results in improvements in overall accessibility of the problem domain application and the evolutionary process. Keywords:Virtual User, visualization, genetic algorithm,domain. 1. Introduction Automated software testing products use scripting languages to describe user behaviors [Bei90,Mer98a,Mer98b]. Executing the scripts emulates users interacting with the test application. Variables within the scripts can change virtual user behaviors. The standard parameterization technique generates a brute force testing method for the application in question. Yet there exist problem domains where brute force is not feasible, including : situations where a variable's range of values is too large; domains where many variables need to be parameterized. While discussing these problems with a consultant in the automated software testing industry, it became clear that their testing techniques could benefit from a genetic algorithm (GA). Instead of a brute force approach for testing, the GA could guide the selection of the parameters used by the virtual users. The GA would search for virtual users that maximized error conditions. GAs could also benefit from taking a virtual user point of view. Since the GA is essentially a generate and test search method, a generated virtual user could test each potential solution. Granted, GAs are applied to many optimization problems where a user interface is ill fitted. But as the popularity of GAs has grown, the domains that GAs are utilized in has grown as well. In these new domains understanding the subtleties of a solution might involve more than just a fitness value and its genetic representation. To understand the ``how'' of a solution one might benefit from a visual interface for the problem domain application. Likewise, the set of users applying GAs is growing more diverse. Communicating the evolutionary processes and advantages becomes more challenging as domains broaden. Often visualization is the most effective aid in fostering an understanding of the underlying phenomena. Recent work [BB98] towards visualizing the evolutionary path greatly aids the understanding of evolutionary processes. This work compliments their approach. This paper describes an approach where each individual in a genetic algorithm is treated as a virtual user. The GA search involves finding the virtual users that manipulate the problem domain application in the most appropriate fashion. True, this paradigm demands more discipline on behalf of the system architect in the early stages of design. However this effort does not go unrewarded, as it leads to

278 256 a system that is easier to test, explore, understand, and share. The payoff is not an optimized GA, but instead a greater accessibility to the problem domain and the results generated by the GA. System architecture An individual in a genetic algorithm acting as a virtual user does not demand a sophisticated scripting system that understands user interfaces. If one uses the Model-View- Controller (MVC) design pattern [Gol84,BMR6#1+96] in the program representing the problem space, an interface is clearly defined that can be utilized by code designed for either a human or virtual user. Figure 1 lays out the basic genetic algorithm as a virtual user manager (GAVUM) architecture. This mechanism leads to relatively painless implementations of virtual users. Figure: A graphical representation of the system architecture. The user's application is composed of the Model, View and Controller. The user affects the Model through the Controller and the Model's state is visualized through the View. The GAVUM mechanism integrates a GA with the problem domain application by designing the GA Individual to use the Model's View and Controller interfaces. In the MVC design pattern, the Model object contains the heart of the problem domain application. It encapsulates all the data and functions underlying the phenomena to model. When something ``interesting'' happens in the Model, it updates all the View objects associated with it. ``Interesting'' is defined by the problem domain and the View itself. An example of these two objects would be a simple model oven. The Model would be tracking the temperature of the oven: calculating heat flow and transfer. As the temperature in the oven changes, it would notify all of its associated View objects with the temperature in degrees Fahrenheit. One View object could be coded to display the temperature, another View object could re-calculate the temperature into Celsius before displaying it, while another View could display the temperature as a color. A Model can have any number of Views, but each View can only have one Model. Adding a Controller object to this application would allow us to alter the oven's behavior. The Model object provides a set of methods through which its behavior and/or state can be changed. The Controller object could map keystrokes and mouse actions in the user interface to the appropriate method calls. The user could be presented with a knob labeled with temperatures. As the knob's value changes, through user actions, the Controller object would alter the settings within the Model. Again, the mapping from Controller to Model is many to one. There could also be a text field where the user can type in a desired temperature. Either way, the change in temperature of the oven would affect the model and in turn, be visualized through the various Views. Once the MVC structure is in place, coding the GA virtual users is straightforward. The process involves these steps: Create a Controller object to map a genetic representation to the appropriate method calls in the Model object. Different genomes would encode for different parameters passed to the Model methods. Create a View object that can interpret the data from the Model and calculate a fitness value. Depending on the way the Model is implemented, fitness values can be derived after one update or after many updates. Glue the Controller and View together in a GA Individual object. This interfaces the GA with the problem application by defining exactly how a fitness value is generated from a genetic representation.

279 257 There is no runtime penalty for visualization for the virtual users, as their data does not need to be displayed. Mutation and crossover operations can act upon the genome in one's preferred fashion. Similarly, one's favorite selection methods can then drive the GA, to evolve the virtual users. This architecture doesn't define any changes to the actual evolutionary computational method. Instead, it improves on the interface among the evolutionary computation, the problem domain application, and the researcher. It also fosters reusability at the level of the GA and the problem domain application. GA Individuals seen as virtual users can control a variety of application types, all being evolved by an unmodified GA. The Model object, which defines the majority of the domain application, is used by both human and virtual users. With one code base, changes to the application are less complicated and consistency is maintained without any effort. 1. Case studies As mentioned earlier, the number of domains that utilize GAs is growing. The domain of multi-agent systems is an example of a new area benefitting from GAs. Multi-agent systems are difficult to design and analyze because the group's behavior often relies on subtle traits in the local behavior of the individuals. GAs allow designers to search through the space of individual behaviors for desired group behaviors. This work presents three different examples of GAs applied to multi-agent systems using the GAVUM architecture. 1.1 Resource Allocation Sen et. al. [SRA96] suggest that limited knowledge in multi-agent systems can be beneficial in certain cases. Their work involves the resource allocation problem, where N autonomous agents must distribute themselves among M resources linked in a circular chain. During a time step each agent chooses a resource to use. The performance of the system can be measured by the number of time steps it takes for the agents to converge upon and maintain the optimal state starting from random initial distributions. In Sen's implementation, each agent's behavior is controlled by identical probabilistic functions. f ii determines the probability of an agent staying at the current resource and f ij defines the probability of moving from resource i to resource j. These are the control functions: Here, r i represents the current number of agents at resource i;,, and are control parameters. Sen introduces a window parameter, which limits the number of resources an agent has knowledge of at any time. This window value defines how many neighboring resources the agent has access to. The window is centered around the agent, so a window value of 3 would allow agents access to the number of agents at the adjacent resources. The window value also defines legal values of j for each i in the control functions. With Sen's control parameters, larger windows led to slower convergence to the optimal state, and smaller windows led to faster convergence. To explain their results, Sen presents the number of agents at one resource over time. As the number of sites visible to the agents increases, the number of agents at the resource varies longer. This counter intuitive phenomena brought forth an attempt to reproduce the results [Bar98a]. First, the Model object was created to represent the resources and the agents. Next, the Controller was designed, creating a user interface through which the control parameters and window value could be set. A View of the resources displays the the number of agents at each resource and the current time step. These components together create the problem domain application. With this program, users can alter the variables and explore the effects the control parameters and window values have on group behavior. It would take one user too long to explore the space of potential control parameters over all the window values. Thus, the GA was introduced, to emulate the actions of many users. The GA's task was to find the virtual user that would create the fastest settling group of agents. The mutation and crossover operations alter the genetic representation which changes the,, and values set by the GA Controller. The GA View object would ignore the updates until the optimal state was reached. Once the optimal state was reached, the current time step was used as a fitness value. In some cases, the optimal state was never maintained. For those cases, the simulation was halted after 30,000 time steps. To gather more accurate statistics, several iterations of the model needed to be run. Additional logic was added to the GA Individuals, enabling them to repeat their experiments in the model. Running the GA found a virtual user with a set of control parameters that provided superior performance regardless of the window values. The GA results were immediately

280 258 examined by plugging the best individual's parameters into the problem domain application. The agents quickly settled into the optimal state as expected. By examining the time series of fitness values, an increase in fitness was apparent. Yet this reveals very little in this domain. But with this system architecture, one can pick out the best individual every 10 time steps and plug the virtual user's values into the stand-alone application. This ends up being much more informative, as the group behavior starts out being very volatile in the early generations. It becomes clear that in these volatile agent situations, the window size has little to do with group behavior. The group performs equally poorly at all window sizes. As the GA individuals evolve, one can observe agent volatility decrease. First for only some window values, then for all the window values. This information gives many hints about the nature of the problem space, allowing the user to generate more hypotheses about the behavior of the system. This example illustrates the GAVUM architecture's ability to enhance the knowledge acquisition process. 1.2 Game Playing Artificial Intelligence Designing artificial intelligence routines for games is another application of the GAVUM architecture. Though the AI in some of today's commercial games might use more information than what is provided to the user, that hasn't always been the case [Sam63]. Additionally, in cognitive science research, AI models are designed with the human player's abilities and techniques in mind [KM92]. These facts support using virtual users to evolve game playing AIs. A simple client/server game of Snake was written as a separate project to learn about networking. Snake is a simple game, where a user controls a snake's movement in a 2-d world. A snake increases in size when food is eaten, and it dies when the snake's head comes in contact with any non-food object in the world. Users interested in playing the game launch a snake View in order to see the world. Launching a snake Controller enables the user to control a snake in the game. Both client programs connect to the snake world Model object, which doubles as the server. With the client-server architecture, multiple snakes can be introduced into the game. Using the GAVUM architecture, a GA individual then can use a snake View (to understand the world and its place in it) and a snake Controller to control its own snake. The GA individual becomes a real-time, interactive player in a networked game of snake. The snake game reports statistics about players, like the longest length attained, number of lives, etc. The GA individual then can use this data to calculate its fitness. Once its fitness is calculated (longest length achieved during 10 games), it quits the snake game. The selection process of the GA decides its fate. This example shows displays how the GAVUM architecture can get co-evolution for free. The networked architecture of the game enables co-evolution in the GA, as the individuals are competing against other snakes controlled by other GA Individuals. Additionally, the GAVUM system even allows visualization of the coevolutionary paths. The user can open up their own snake View of the game, in order to watch the GAVUM snakes play and evolve. The user can also join the game, affecting the fitness of the individuals in the GA. Many benefits are reaped simply by using a different program structure. This architecture could be directly applied to the iterated prisoner's dilemma domain as well as its related tournament models. 1.3 Artificial Life Models My research focuses on artificial life (AL) models. These models rely on visualization to express their behavior. Again, through the model's API designed for use with the Views and Controllers, the GA individual has control over the Model, without having to compute the display. My models are designed to facilitate examining the evolution of coordination in multi-agent systems [Bar98b]. In my model, survival is the task facing homogenous populations of agents. The model's agent environment has many parameters that can be adjusted, offering a variety of ways to increase its hostility. The behavior of the agents themselves can be adjusted via the Controller. The agent's behavior is what the GA virtual users adapt. Each GA run evolves agents best fit for a given environment. In these experiments, GA individuals are created that define new environmental conditions in which to evolve agents. The researcher now takes the position of defining meta-experiments, with the virtual users performing sets of experiments. This allows one to study the environmental effects on the agents and their evolution. Each GA run creates an evolutionary path that represents a set of experiments, each resulting in agents performing to varying degrees of success. And then collectively, the GA runs define another set of evolutionary paths across meta-experiments. Hopefully, the automation of the experiments that is intrinsic in the GAVUM architecture will increase the rigor of AL research. As I extend the model to allow heterogeneous populations, the same GA will be used, with some modifications to the GA Individual object. Again, coevolution will be a convenient side effect.

281 259 This AL domain highlights another benefit of this paradigm. With the other models, reproducibility was not as much of an issue, because the code was simpler and there were fewer variables in play. But with artificial life models, programmers are only limited by their imagination and the programs quickly get rather complex. Reproducibility has been a problem facing the AL community from the beginning. By encouraging the MVC and virtual user system, programmers might find it worthwhile to introduce useable Views and Controllers. Then the problem domain applications can stand-alone and be shared with other researchers. Instead of trying to duplicate the code, others can use the same code to repeat the experiments and run their own new ones. 2. Discussion All too often, researchers are too excited to start evolving individuals in a GA, and not enough effort is put into figuring out how to explain the results. The GAVUM architecture provides a step towards insuring that the most can be learned from the evolutionary computation. Diligent application of the GAVUM paradigm will also allow others to use the problem domain application and to explore the problem domain themselves. With the Java programming language and applets that can be run through web browsers, accessibility of the problem domain application can become trivial. URLs of all work discussed here is soon to come. The GAVUM architecture is currently being discussed with instructors of an artificial intelligence class as a way of introducing the class to GAs. First, students will be introduced to modelling through a particular problem domain application. Through hands on experience with the application, the students gain a deeper understanding of the domain problems that the model represents. Then, as they become familiar with the program, they would be asked to create their own metric for comparing two instances of their models with each other. In essence, each student is implicitly creating a fitness function. Next, the students would be introduced to a method of finding individuals that maximize performance according to a metric - the GA. They can then code up their own GA individual. The students would already be familiar with the kinds of information expected by the Controller, as they have had to supply the same information to the program when they were using it. Through creating their own metric for comparison, they have used information from the View. The students just have to formalize this information in actual code. This clearly defines assignments in the sense of coding routines with well documented input and output. Additionally, these methods probably would not be too technically demanding, which can be a benefit in introductory classes. But the elements of the model that are tested, how the individuals are evaluated, and the genetic representations are completely open ended. Each student has the freedom to explore the areas of the model that are personally interesting. After running the experiments, (most likely editing their GA individuals along the way), the students can then discuss the various advantages and disadvantages of their different representations, fitness functions, and evolutionary techniques (mutation and crossover operators and rates, plus the selection methods). Throughout the assignments, the process reinforces the benefits of the scientific method. Hopefully the advantages of the GAVUM architecture are apparent and seem worth the design and coding effort. Programming as if people mattered [Bor91] is a honest, practical book that stresses the need to bridge the gap between programmers and users. The same need is found in the GA community. Although GAs came from a programmer-rich environment, one can no longer assume that users of GAs will only be other programmers, due to the growing diversity in the GA user population. Both the GA and the problem domain application should be considered tools to be used by others for learning - making visualization all the more important. Currently, work is being conducted to integrate the GA into the actual problem domain application. User interface issues need to be refined in order to merge the View and Controller of both the GA and the problem domain application. Future work is directed towards better formalizations of the model. There are most likely better ways to generalize the API in order to create more reuseable code. Finally, applying this approach to other domains would help evaluate the effectiveness of the GAVUM architecture Acknowledgements I am grateful to Catherine Kim at Mercury Interactive for numerous discussions about automated software testing, virtual users, and their potential relationships with genetic algorithms. This research is supported by the National Science Foundation under grants GER and CDA References [1]. Baray.Effects of individual decision schemes on group behavior. In The Third International Conference on Multi-Agent Systems, New York, NY, IEEE Press.

282 260 [2].Baray. Effects of population size upon emergent group behavior. In Complex Systems '98 Conference Proceedings, [3].M.A. Bedau and C.T. Brown. Visualizing evolutionary activity of genotypes. Technical Report , The Santa Fe Institute, Santa Fe, NM, [4].Beizer. Software Testing Techniques. International Thomson Publishing, Boston, MA, [5].Frank Buschmann, Regine Meunier, Hans Rohnert, Peter Sommerlad, and Michael Stal. Pattern Oriented Software Architecture: A System of Patterns. John Wiley and Sons, New York, NY, First Author: Msc (computer Science) On 2006, M.Tech (Computer Technology) On 2010,National Institute of Technology, Raipur, CG Two international paper, Generally interest in wireless, image processing and genetic algorithm. Second Author Msc (Mathematics), MCA (Computer Application), National Institute of Technology, Raipur, CG Two international paper, Generally interest in wireless, image processing and genetic algorithm. Third Author BSC (Electronics), MCA (Computer Application),National Institute of Technology, Raipur, CG Two international paper, Generally interest in wireless, image processing and genetic algorithm. [6]. N.Borenstein. Programming as if people mattered. Princeton University Press, Princeton, NJ, [7].Adele Goldberg. Smalltalk-80: the interactive programming environment. Addison-Wesley, Reading, MA, [8].D.Kirsh and P. Maglio. Reaction and reflection in tetris. In J. Hendler, editor, Artificial intelligence planning systems: Proceedings of the First Annual International Conference (AIPS92), San Mateo, CA, Morgan Kaufman. [9].Mercury Interactive. Load Runner : Controller Users Guide. Mercury Interactive Press, [10].Mercury Interactive. Load Runner : Creating Vuser Scripts. Mercury Interactive Press, [11].A.L.Samuel. Some studies in machine learning using the game of checkers. In E.A. Feigenbaum and J. Feldman, editors, Computers and Thought. McGraw-Hill, New York, NY, [12].S. Sen, S. Roychowdhury, and N. Arora. Effects of local information on group behavior. In Mario Tokoro, editor, The Second International Conference on Multi-Agent Systems, Menlo Park, CA, AAAI Press.

283 261 Performance Analysis of Peak-to-Average Power Ratio Reduction Techniques for Wireless Communication Using OFDM Signals Pawan Sharma 1, Seema Verma 2 1 Department of Electronics and Communication, GGSIP University, Bhagwan Parshuram Institute of Technology, Delhi , India 2 Department of Electronics, Banasthali University Rajasthan , India Abstract Orthogonal Frequency Division Multiplexing (OFDM) has been currently under intense research for broadband wireless transmission due to its robustness against multipath fading. However OFDM signals have a problem with high Peak-to- Average power ratio (PAPR) and thus, a power amplifier must be carefully manufactured to have a linear input-output characteristic or to have a large input power back-off. In this paper, some of the important PAPR reduction techniques which have been compared based on computational complexity, bandwidth expansion, spectral spillage and performance. Keywords: Orthogonal Frequency Division Multiplexing (OFDM), Peak-to-Average Power Ratio (PAPR), Complementary Cummulative Distribution Function (CCDF), High Power Amplifier (HPA), Partial Transmit Sequence (PTS), Selected Mapping (SLM). 1. Introduction The growth of mobile wireless communications has been producing the demand for high-speed, efficient and reliable communication over the hostile wireless medium. As a modulation scheme for such applications, Orthogonal Frequency Division Multiplexing (OFDM) possesses several desirable attributes, such as immunity to the inter-symbol interference, robustness with respect to multi-path fading, and ability for high data rates. Thus, OFDM has been proposed in various wireless communication standards such as IEEE802.11a standard for wireless Local Area Networks (WLAN), IEEE802.16a standard for Wireless Metropolitan Area Networks (WMAN), digital audio/video broadcasting, Terrestrial Digital Video Broadcasting (DVB-T), the ETS1 HIPERLAN/2 standard and high speed cellular data [1]. However, one of the major drawbacks of OFDM system has been its high Peak-to-Average Power Ratio (PAPR). The high PAPR brings the OFDM signal distortion in the non-linear region of high power amplifier (HPA) and the signal distortion induces the degradation of bit error rate (BER). Moreover, to prevent spectral growth of the multicarrier signal in the form of intermodulation among subcarriers and out-of-band radiation, the transmit power amplifier has to be operated in its linear region [2]. If the HPA is not operated in linear region with large power back-offs, it is impossible to keep the out-ofband power below the specified limits. This situation leads to very inefficient amplification and expensive transmitters. Therefore, it has been important and necessary to research on the characteristics of the PAPR, including its distribution and reduction, in OFDM systems, in order to utilize the technical features of the OFDM. In 2005, the Wimedia alliance worked with the European Computer Manufactures Associations (ECMA) and announced the establishment of the WiMedia MB-OFDM (Multiband Orthogonal Frequency Division Multiplexing), Ultra- Wideband (UWB) radio platform as their global UWB standard, ECMA-368 and the latest updated version incorporated spectral nulling. ECMA -368 was also chosen as the physical layer (PHY) for the high data rate wireless specifications, for high-speed wireless USB (W-USB), Bluetooth 3.0 and wireless high-definition Media Interface (HDMI) [3]. To reduce the PAPR several techniques have been proposed such as partial transmit sequences (PTS) [4, 5], selective mapping (SLM) [6, 7], clipping [8, 9] clipping and filtering [10], coding [11], tone reservation (TR) and tone injection (TI) [12]. Each of these methods has a different cost for the reduced PAPR. Although some techniques of PAPR reduction have been summarized, it is still necessary to give a comprehensive review of PAPR reductions in terms of transmission power, data rate loss, implementation complexity and BER performance, etc.

284 2. OFDM SYSTEM MODEL In OFDM system, a block of N symbols, X n,n = 1, 2...N-1, is formed with each symbol modulating one of a set of subcarriers, { f n, n = 0, 1, 2. N1}, where N is number of subcarriers. The N subcarriers are chosen to be orthogonal, that is, f n = nf = n/t, where T is the original symbol period. The OFDM baseband signal is denoted by: Where j = 1. A guard Interval is inserted to each OFDM symbol as a cyclic prefix. Finally, an OFDM signal stream is obtained after demultiplexing. The PAPR of OFDM is defined as, 0 (2) Where E{ } is the expectation. In principle, there has been more concern with reducing the PAPR of the continuous-time OFDM signals, since the cost and power dissipation of the analog components often dominates. However, most existing PAPR reduction methods can only be implemented on the discrete-time OFDM signals by IFFT. The IFFT output can be expressed as follow: keep a low BER, it requires a linear work in its linear amplifier region with a large dynamic range. However, this linear amplifier has poor efficiency and is so expensive. Power efficiency is very necessary in wireless communication as it provides adequate area coverage, saves power consumption and allows small size terminals etc. It has been therefore important to aim at a power efficient operation of the non-linear HPA with low back-off values and try to provide possible solutions to the interference problem brought about. Hence, a better solution has been to try and prevent the occurrence of such interference by reducing the PAPR of the transmitted signal with some manipulations of the OFDM signal itself. Large PAPR also demands the Digital-to-Analog Converter (DAC) with enough dynamic range to accommodate the large peaks of the OFDM signals. Although, a high precision DAC supports high PAPR with a reasonable amount of quantization noise, but it might be very expensive for a given sampling rate of the system. Whereas, a lowprecision DAC would be cheaper, but its quantization noise would be significant, resulting in reduction of the Signal-to- Noise Ratio (SNR) when the dynamic range of DAC is increased to support high PAPR. Furthermore, OFDM signals show Gaussian distribution for large number of subcarriers, which means the peak signal rarely occur and uniform quantization by the ADCs is not desirable. If clipped, it will introduce in-band distortion and out-of-band radiation (adjacent channel interference) into the communication systems. Therefore, the best solution has been to reduce the PAPR before OFDM signals have been transmitted into nonlinear HPA and DAC. 3. LATEST PAPR REDUCTION TECHNIQUES IN OFDM SYSTEMS 262 the corresponding PAPR can be defined as 2.1 Motivation Of PAPR Reduction Most radio systems employ the HPA in the transmitter to obtain sufficient transmission power. For the purpose of achieving the maximum output power efficiency, the HPA is usually operated at or near the saturation region. Moreover, the nonlinear characteristic of the HPA is very sensitive to the variation in signal amplitudes. However, the variation of OFDM signal amplitudes is very wide with high PAPR. Therefore, HPA will introduce inter-modulation between the different subcarriers and introduce additional interference into the systems due to high PAPR of OFDM signal. This additional interference leads to an increase in BER. In order to lessen the signal distortion and (4) In this section, various techniques for reducing the PAPR have been discussed. 3.1 Clipping And Filtering The OFDM signal contains high peaks so Clipping And Filtering (CAF) is used in this system [13]. The generated OFDM signal is transferred to the clipping block. In the clipping part, when amplitude exceeds a threshold, the amplitude is hard-clipped while the phase is saved.

285 263 Fig.1. The OFDM transmitter including clipping scheme. Namely, when we assume a phase of baseband OFDM signal is and the threshold is A, the output signal after clipping is shown as There have been four other clipping techniques: Classical Clipping (CC), Heavyside Clipping (HS), Deep Clipping (DC) and Smooth Clipping (SC) whose functions are depicted in Fig.2. [15]. Rewriting the discrete-time OFDM signal to polar coordinates gives = r n e jn,where r n represents the amplitude of and n represents the phase of The clipped signal is expressed as (6) Where the clipping function f( r ) is expressed below according to the type of clipping used. The clipping causes both in-band and out-of-band distortion because of non-linear operation of the clipping, The in-band distortion causes degradation of BER, while distortion also causes out-of-band emission. To decrease the interference to neighboring channels, out-of-band components must be reduced with a band-limiting filter. Due to the non-linearity, clipping operations causes out-of-band components. Suppressing the components by band-limiting filters makes clipped peak signal expand. When CAF is applied to an oversampled signal, the peak regrowth becomes comparatively small [14]. However, it has been inevitable that the PAR is larger than the clipping threshold. For these reasons, some peak regrowth reduction methods have been proposed. A straightforward way of peak growth reduction has been repeated CAF method. In this method, clipping and filtering are applied repeatedly, and peak regrowth is reduced gradually. In the paper it has been shown, that repeated clipping and filtering significantly reduces the PAR. The method is effective, however, the calculation cost increases proportionally to the number of repeats, and filtering delay also increases. So there has been a tradeoff between PAR and the system cost. Fig.3. PAPR reduction performance of CC,HC,HC and DC techniques. The Classical Clipping (CC) proposed in [16] has been one of the most popular clipping techniques for PAPR reduction known in the literature [11]. It is sometimes called hard clipping or soft clipping, to avoid any confusion; it is called Classical Clipping (CC) in this paper. The function-based clipping used for CC technique is defined below and depicted in Fig. 2(a)., where A is the clipping level. (7) Heavyside Clipping, often called hard clipping has been used in [17] as a baseband nonlinear transformation technique to improve the overall communication system performance. The heavyside function is expressed below and depicted in Fig.2(b). f ( r) = A, r 0 (8) Fig.2. Functions-Based Clipping for PAPR Reduction Deep Clipping (DC) has been proposed in [14] to solve the peaks regrowth problem due to the out-of-band filtering. So, in DC technique, the clipping function has been modified in order to deeply clip the high amplitude peaks. A parameter called clipping depth factor has been introduced in order to control

286 the depth of the clipping. The function-based clipping used for DC technique is defined below and depicted in Fig.2 ( c ). 264 Fig.5. Block Diagram of PTS Technique Therefore, Fig.4. BER performance of CC,SC,HC and DC techniques. The function for Smooth Clipping (SC) [18] has been defined below and depicted in Fig. 2(d ),where b= (10) 3.2 Partial Transmit Sequences (PTS) For PAPR reduction using partial transmit sequence a typical OFDM system with input data block in X has been partitioned into M disjoint subblocks of clusters, which are represented by the vectors {X (m), m = 0,1, M1}[19]. Fig.5. shows the block diagram of the PTS technique. Where X (m) (m) = [X 0 X (m) 1. X (m) N1 ] with X (m) k = X k or 0 (0 m M 1). In general, for PTS scheme, the known subblock partioning methods can be classified into three categories: adjacent partition, interleaved partition and pseudo-random partition. Then, the subblocks X (m) are transformed into M time-domain partial transmit sequences x (m) = [ ] = IFFT LNxN [X (m) ] (12) These partial sequences are independently rotated by phase factors b = {b m = e jm, m = 0,1, M1}. The objective has been to optimally combine the M subblocks to obtain the time domain OFDM signals with the lowest PAPR Assuming that there are W phase angles to be allowed, thus b m has the possibility of W different values. Therefore, there are W M alternative representations for an OFDM symbol. The PTS technique significantly reduces the PAPR, but unfortunately, finding the optimal phase factors has been a highly complex problem. In order to reduce the search complexity, the selection of the phase factors has been limited to a set of finite number of elements. The Exhaustive Search Algorithm (ESA) [20] has been employed to find the best phase factor. However, the ESA requires an exhaustive search over all combinations of the allowed phase factors and has exponential search complexity with the number of subblocks. To reduce the computational complexity, some simplified search techniques have been proposed such as the Iterative Flipping Algorithm (IFA) [21]. Although the IFA significantly

287 reduces the search complexity, there has been some gap between its PAPR reduction performance and that of the ESA. A Cross-Entropy (CE) based method has been proposed by Jung-Cheih Chen [20] for obtaining the optimal phase factors for the PTS technique to reduce the PAPR. Jung-Cheih Chen [22] has proposed a Quantum-Inspired Evolutionary Algorithm (QEA) based method to obtain the optimal phase factor for the PTS technique. Abolfazl et al.[23].has proposed an Auto- Correlation Function (ACF) to develop a new PTS subblocking technique using Error-Correcting Codes (ECCs). This technique minimizes the number of repeated subcarrier with a subblock and provides better PAPR reduction than pseudorandom or m-sequence subblocking. Fig.6.PAPR performance of an OFDM signal 3.3 Selected Mapping Technique In SLM, the input data sequences have been multiplied by each of the phase sequences to generate alternative input symbol sequences. Each of the alternative input data sequences is made the IFFT operation, and then the one with the lowest PAPR is selected for transmission [11]. A block diagram of the SLM technique has been depicted in Fig.7. Each data block is multiplied by V different phase factors, each of length N, B v = [ b,0, b,1,..b u, N-1 ] T ( = 0,1,..V-1), resulting in V different data blocks. Thus, the v th phase sequence after multiplied is X = [X 0 b,0, X 1 b, 1,, X N1 b,n1 ] T ( = 0,1,..V-1). Therefore, OFDM signals becomes as Where 0 t NT, = 1,2,,V1. Fig.7. Block Diagram of SLM Technique Among the data block X ( = 0,1,.V-1), only one with the lowest PAPR has been selected for transmission and the corresponding selected phase factors b,n also should be transmitted to receiver as side information. For implementation of SLM OFDM systems, the SLM technique needs V IFFT operation and the number of required bits as side information is [log 2 V] for each data block. Therefore, the ability of PAPR reduction in SLM depends on the number of phase factors V and the design of the phase factors. Some extension of SLM also has been proposed to reduce the computational complexity and number of the bits for side information transmission [24]. In symbol scrambling techniques the input data sequence has been scrambled using a number of specialized scrambling sequences. The sequence which produces the lowest PAPR is the one used for transmission. Selected Mapping (SLM) has been one of the most popular signal scrambling techniques used to reduce the PAPR of OFDM signals. The SLM technique proposed by Bauml et al. [26] takes the OFDM subcarrier data block to be transmitted and multiples it element-wise by a number of phase adjustment vector sets. The new statistically independent phase adjusted OFDM frames represent the same transmitted information, but have different PAPR values. The OFDM frame that has the lowest PAPR is then selected to be transmitted. As reported in [26] the SLM technique can provide a 0.1% probability PAPR reduction of 2.8 db (from 10.4 db to 7.6 db) when applied to 128-subcarrier QPSK-OFDM symbols. However, one drawback of this technique is that some additional information relating to the phase vector set that produces the lowest PAPR also requires to be transmitted along with the OFDM signal. This extra information increases the overhead. Breiling et al.[25] have proposed an SLM technique that avoids explicit use of side information by appending a label that lets the receiver identify the scrambled sequence which produces the lowest PAPR. This scrambler uses a number of shift-register stages with internal feedback paths resulting in error propagation in the receiver causing an increased reception bit-error-rate (BER). Various SLM techniques without side information have been proposed, e.g., Jayalath and Tellambura [27] have proposed a maximum likelihood (ML) decoder for SLM and Partial Transmit Sequences (PTS) techniques that requires no side information. The receiver exploits the large Hamming distances in the set of 265

288 the phasing sequences and uses optimum hard decision for each subcarrier. Stephane and Samer [28] have proposed a SLM technique without side information. This method takes into account the increase in average energy with PAPR reduction which is identical to that of classical SLM. A recursive Selected Mapping (RSLM) for PAPR reduction has been proposed by Lingyin and Yewen [29]. where the PAPR reduction has been a little better than that of SLM with high reduction in computational complexity. Xiaowen and Seungmin [30] have proposed a look-up table method (LUT) by making use of the feature that the PAPR performance is independent of modulation schemes in normal OFDM. This method shows the regulations of selective efficient phase rotation factors. It proves a way of achieving the most efficient PAPR performance for SLM-OFDM, but limited to FFT size 8. Fig.8. CCDF of the PAPR obtained with the SLM techniques. In TR, the objective has been to find the time domain signal c to be added to the original time domain signal to reduce the PAPR. Let {c= denote complex symbols for tone reservation at reserved tones. Thus, the data vector changes to x + c after tone reservation processing, and this results in a new modulated OFDM signals as Where C = IFFT(c). Therefore, the main aim of the TR has been to find out the proper c to make the vector with low PAPR. To find the value of c, a convex optimization problem must be solved which can be easily cast as a linear programming problem. Similarly, TI also uses an additive correction to optimize C in (15). The basic idea of TI has been to extend the constellation and thus the same data point corresponds to multiple possible constellation points. One option has been to replicate the original shaded constellation into several alternative ones. Therefore, C has been a translation vector such that C = ( ). Note that TI needs not require the extra side information and the receiver only needs to know how to map the redundant constellations on the original one. Some modifications of TI have been proposed to obtain good performance including PAPR reduction and low complexity [31]. S.Janaththanan et.al [ 32] have proposed a novel gradient based approach for PAPR reduction using TR technique. The algorithm performs better but depends on the pilot locations. The TI technique has been more problematic than the TR technique since the injected signal occupies the frequency band as the information bearing signals. Moreover, the alternative constellation points in TI technique have an increased energy and the implementation complexity increases for the computation the optimal translation vector Tone Reservation and Tone Injection Tone Reservation (TR) and Tone Injection (TI) are two efficient techniques to reduce the PAPR of OFDM signals [11]. Fig.9. describes the block diagram of TR and TI, in which the key idea is that both transmitter and receiver reserve a subset of tones for generating PAPR reduction signals, c. Note that these tones are not used for data transmission. Fig.9. Block Diagram of TR/TI approaches for PAPR Reduction. 3.5 CodingTechniques Coding can also be used to reduce the PAPR. A recursive convolutional code, has been often used in OFDM system as a channel code. Yan Xin et al. [33] have proposed an integration of guided scrambling (GS) coding with SLM and PTS. The new GS-SLM and GS-PTS systems do not require transmission of side information and can be implemented resulting in good PAPR performance. Reed-Solomon and simplex codes has been proposed by Robert et al. [34] for PAR reduction in OFDM. An adaptive coding technique has been proposed by Zafar et al. [35] for reducing PAPR in COFDM to achieve reduction in PAPR as well as error correction capability. The adaptive approach has been adopted in order to reduce hardware for a slight increase in complexity. The coding technique Error Correction (EC) SLM using convolutional codes proposed by K.Khan et.al[36] has been studied. Guided polynomial scrambling technique gives the

289 most effective PAPR performance but slight degradation of BER. 4. CONCLUSIONS In this paper some PAPR reduction techniques for multicarrier transmission have been discussed. Many techniques to reduce the PAPR have been proposed all of which have the potential to provide substantial reduction in PAPR at the cost of loss in data rate, transmit signal power increase, BER increase, computational complexity increase and so on. No specific PAPR reduction technique has been the best solution for all multicarrier transmission system. It has been suggested that the PAPR reduction technique should be carefully chosen according to various system requirements. References [1] S.H.Han and J.H.Lee, An Overview of Peak-to-Average Power Ratio Reduction Techniques For Multicarrier Transmission, IEEE Wireless Communications,Vol.12,No.2, Apr.2005, pp [2] A. Ghassemi and T. Aaron Gulliver, A low-complexity PTS-Based Radix FFT Method for PAPR Reduction in OFDM Systems, IEEE Transactions on Signal Processing, Vol.56, No.3, March 2008, pp [3] R.S.Sheratt and O.Cadenas. J, A Double Data Rate Architecture for OFDM Based Wireless Consumer Devices, IEEE Transactions on Consumer Electronics, Vol.56, No.1, February 2010, pp [4] S.H.Han and J.H.Lee, PAPR Reduction of OFDM Signals using a Reduced Complexity PTS Technique, IEEE Signal Processing Letters, Vol.11, No.11, November 2004, pp [5] T.Jiang, W.Xiang, P.C.Richardson, J.Guo, and G. Zhu, PAPR Reduction of OFDM Signals Using Partial Transmit Sequences with Low Computational Complexity, IEEE Transactions on Broadcasting, Vol.53, No.3, September 2007, pp [6] C.L.Wang, M.Y.Hsu, and Y.Wuynag, A Low-Complexity Peak-to- Average Power Ratio Reduction Technique for OFDM systems, Global Telecommunication Conference,2003 Globecom 03 IEEE, pp [7] Y.C.Cho, S.H.Han and J.H.Lee, Selected Mapping Technique with Novel Phase Sequences for PAPR Reduction of an OFDM Signal, in Proc. of 5th IEEE VTC 2004-Fall, vol. 7, Sep. 2004, pp [8] G.Hill and M.Faulkner, Comparison of Low Complexity Clipping Algorithms for OFDM, IEEE International Symposium on Personal, Indoor and Mobile Radio communications, 2002,Vol..1, pp [9] H.Ochiai and H.Imai, Performance Analysis of deliberately clipped OFDM signals, IEEE Trans. on communications, Vol.50, No.1, January 2002, pp [10] S.KYusof and N.Fisal, Coorelative Coding with Clipping and Filtering Technique in OFDM Systems, ICICS-PCM 2003,Singapore, IEEE 2003, pp [11] T.Jiang and Y.Wu, An Overview: Peak-to-Average Power Ratio Reduction Techniques for OFDM Signals, IEEE Transactions on Broadcasting, Vol.54, No.2, June 2008, pp [12] D.Guel and J.Palicot, FFT/IFFT Pair based Digital Filtering for the Transformation of Adding Signal PAPR Reduction Techniques in Tone Reservation Techniques, Fifth International Conference on Wireless and Mobile communications, (ICWMC 2009), August [13] K.D.Rao and T.S.N.Murthy, Analysis of Effects of Clipping and Filtering on the Performance of MB-OFDM UWB Signals, Proc. of the th International Conference on Digital Signal Processing (DSP 2007), IEEE, pp [14] S.Kimura,T.Nakamura,M.Saito and M.Okada, PAR reduction of OFDM signals based on deep clipping, ISCCSP 2008, Malta, March 2008, IEEE, pp [15] D.Guel and J.Palicot, Analysis and Comparison of Clipping Techniques for OFDM Peak-to-Average Power Ratio Reduction, International Conference on Digital Signal Processing (DSP 2009), IEEE. [16] X.Li and L.J.Cimini, Effects of Clipping and Filtering on the Performance of OFDM, IEEE Communication Letters, Vol.2, No.5, May 1998, pp [17] Q.Hua,R.Raich,G.T.Zhou, On the Benefits of Deliberately Introduced Baseband Nonlinearities in Communication Systems,, IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 04), Vol.2, May 2004, pp [18] P.Boonsrimuang,E.Puttawong, H.Kobayashi and T.Paungma, PAPR Reduction Using Smooth Clipping in OFDM Systems, The 3rd Information and Computer Engineering Postgraduate Workshop 2003 (ICEP 2003), Jan2003, pp [19] T.Ya-fei, D.Rong-hua, Y.Xiao-an and T.Hai-wei, PAPR Reduction of OFDM Signals Using Modified Partial Transmit Sequences, 2 nd International Conference on Image and Signal Processing, 2009, CISP 09, pp.1-4. [20] J.C.Chen, Partial Transmit Sequences for Peak-to-Average Power Ratio Reduction of OFDM Signals With the Cross-Entropy Method, IEEE Signal Processing Letters, Vol.16, No.6, June 2009, pp [21] L.Yang, R.S.Chen, Y.M.Siu, and K.K.Soo, PAPR Reduction of an OFDM Signal by use of PTS with Low Computational Complexity, IEEE Trans. Broadcast., Vol.52, No.1, March 2006, pp [22] J.C.Chen, Application of Quantum-Inspired Evolutionary Algorithm to Reduce PAPR of an OFDM signal Using Partial Transmit Sequences Technique, IEEE Trans. On Broadcasting, Vol.56, No.1, March 2010, pp [23] A.Ghassemi and T.A. Gulliver, PAPR Application of OFDM Using PTS and Error-Correcting Code Subblocking, IEEE Trans. On Wireless Communications, Vol.9, No.3, March 2010, pp [24] S.H.Han and J.H.Lee, PAPR Reduction of OFDM Signals using a Reduced Complexity PTS Technique, IEEE Signal Processing Letters, Vol.11, No.11, Nov.2004, pp [25] H.Breiling,S.H.Muller-Weinfurtner,and J.B.Huber, SLM Peak-Power Reduction without Explicit Side Information, IEEE Communication Letters, Vol.5, No.5, June 2001, pp [26] R.W.Bauml, R.F.H.Fisher and J.B.Huber, Reducing the Peak-to- Average Power Ratio of Multicarrier Modulation by Selected Mapping, IEEE Electronic Letters, Vol.32, Oct. 1996, pp [27] A.D.S.Jayalath and C.Tellambura, SLM and PTS Peak-Power Reduction of OFDM Signals without side information, IEEE Transactions on Wireless Communications, Vol.4, Sept.2005, pp [28] S.Y.Le Goff, Al-Samahi, S.S.B.K.Khoo, C.C.Tsimenidis and B.S.Sharif, Selected Mapping without side information for PAPR Reduction in OFDM, IEEE Transactions on Wireless Communications, Vol.8, No.7, July. 2009, pp [29] L.Wang and Y.Cao, Improved SLM for PAPR Reduction in OFDM Systems, International Workshop on Intelligent Systems and Applications, 2009, ISA2009, pp.1-4. [30] X.Gu, S.Baek, and S.Park, PAPR Reduction of OFDM Signal Using an Efficient SLM Technique, ICACT 2010, Feb.2010, pp [31] J.Armstrong, Peak-to-Average Reduction for OFDM by Repeated Clipping and Frequency Domain Filtering, IEEE Electronics Letters, Vol.38, No.5, May 2002, pp [32] S.Janaaththanan,C.Kasparis and B.G.Evans, A Gradient Based Algoritm for PAPR Reduction of OFDM using Tone Reservation Technique, Vehicular Technology Conference,2008,VTC Spring 2008, pp [33] Y.Xin and I.Fair, Peak-to-Average Power Ratio of an OFDM signal using Guided Scrambling Coding, Global Telecommunication conference,globecom 2003,IEEE, Vol.4, pp

290 [34] ] R.F.H.fischer and C.Siegl, Reed-Solomon Solomon and Simplex Codes for Peak-to-Average Power Ratio reduction in OFDM, IEEE trans.on Information Theory, Vol.55,No.4, Apr.2009, pp [35] Z.Q.Taha and X.Liu, An Adaptive Coding Technique for PAPR Reduction, Global Telecommunication conference, Globecom 2007, IEEE, pp [36] K.Khan and S.A.Sheikh, PAPR Reduction of OFDM Signals using Convolutional Codes, Proc. of 2009 IEEE Student Conf. on Research and Development, Nov 2009, UPM Serdang, Malaysia, pp Pawan Sharma received the B.E. degree in 1995 and M.E. degree in 2004 from Delhi College of Engineering in Electronics and Communication. Worked in Bharati Vidyapeeth s College of Engineering, Amity Institute of Telecommunication. At present working in Bhagwan Parshuram Institute of Technology as an Assistant Professor in Electronics and Communication department. (Dr.) Seema Verma obtained her Ph.D degree from Banasthali University in She is working as Associate Professor [Electronics] at Banasthali University. She is an active research supervisor and has presented many papers in various international conferences. She has published many research papers in various journals of repute. Her research areas are Coding theory, TURBO Codes, Wireless sensor networks, Network Security & VLSI Design.

291 269 To Design Voice Control Keyboard System using Speech Application Programming Interface Md. Sipon Miah 1, and Tapan Kumar Godder 2 1 Department of Information and Communication Engineering, Islamic University, Kushtia, 7003, Bangladesh. 2 Department of Information and Communication Engineering, Islamic University, Kushtia, 7003, Bangladesh. Abstract With the passing of days men are more dependent on electronic devices. The Main objective of this project is to design and develop a voice Control Keyboard Systems, fully controlled by a computer, and display output on the display device with predefined time. So this project will work as a helping system for those person who has small knowledge about computer system even those person who are illiterate they can operate computer system. We can implement this developed system in other system for example voice control car system. This project voice is the input, sound application programming interface(sapi 4.1) recognize this voice transfer the command to the microprocessor according to the programming code and display device displays the output. Keywords: PC with Pentium microprocessor, a microphone, HMM and sound application software (SAPI). 1. Introduction Day by day our life becomes busy. We have to do a lot of work everyday. For these purpose we use many kinds of computer control system. If we control the Computer system using voice then we can save enough time to do other sophisticated work. By doing this we can makes our work easier and faster. There are several different possibilities for environment control for physically disabled persons. In many cases speech is the most convenient and easy-tolearn alternative. In this study, we wanted to explore the possibility of voice control without being involved in major hardware system development. Thus, our guiding principle was that the system should connect easily to standard products for environmental control, be expandable and use standard equipment to the largest possible extent. The person that volunteered to try out this application already had a sip-and-puff operated system for limited environmental control through infra-red (IR) techniques. Rather than replacing this system, we decided to expand it with voice control and new possibilities. In this way, we could use the old system as back-up in case of system break-downs and we also got an immediate evaluation of relative drawbacks and merits of the two control modes. In developed countries most of the sectors are computerized but in Bangladesh, computer is used mostly in education, office work and printing purpose. If we can control our computer using voice, we lead a smart and faster life. So we have to try to use this advantage. Because the Voice-control is trained to recognize the individual voice pattern, operation by a third person is only possible by means of command keys. Speaking a command is the same as selecting a command with the scroll key and confirming it with the OK key on the pilot or on an external keyboard. By connecting special operating peripherals ( suck/blow switch, foot switch, head switch,...) to the keyboard interface, the Voicecontrol can also be operated by people with a speech impediment. The Voice-control is delivered together with configuration software, allowing the Voice-control to be adjusted to suit the individual. The software can be run on a standard PC under MS-Windows. (CPU Intel or higher, RAM 4 MB, MS-Windows V3.1 or higher) The selected commands are trained on the PC with the help of the Voice-control. Each word is spoken several times so that a common voice pattern can be analyzed. The voice pattern of the word is part of a neural network that allows speech to be recognized in operation. The voice patterns are stored in the pilot. Individual words can be retrained on the Voice-control even without a PC and configuration software. 2. Voice Control System Voice-controlled systems have become more and more popular in the last few years, but in many cases the focus has been on what is technically possible to do, more than what people really want from their systems. We will focus on the human aspect, and try to figure out what the most intuitive way of communicating with a voice-

292 270 controlled system is. We are also interested in finding out how people adapt the way they talk when they are talking to a computer. Perhaps people want a shorter reply from the system than in other cases and also want to express themselves in shorter sentences? In order to build human friendly systems in the future we need to find out how people want their systems to work and perform instead of building more and more technically advanced systems that nobody asked for. We want to explore how different people talk when they talk to a computer compared with when they talk to a human. How does the communication differ with respect to syntax, pragmatics, phonetics and semantics, depending on whether they talk to a computer system or a human? Does a voice-controlled computer system have to be able to handle everything that a human can understand? Do people talk with longer sentences or do they choose to give short commands? What seems to be the intuitive way to speak? focus on voice control would be to find out how people really talk with the systems instead of how the systems work. 2.1 Dialogue Systems A dialogue is a spoken, typed or written interaction in natural language between two or more agents. It is divided into units called turns, in which a single speaker has temporary control of the dialogue and speaks/writes for some period of time. Within a turn, the speaker may produce several spoken or typed utterances1. A turn can consist of more than one move. A move is a speech act in the sense that it is an act that a speaker performs when making an utterance, such as questions, warnings, statements etc2. It has a functional relation to the conversation of which it is a part. A dialogue system is a system that allows a human, the user, to use natural language in the interaction with the computer. In the same way, the computer replies with natural language. The natural language can be either spoken or written, either complete sentences or fragments of sentences. An example of a dialogue system is the SJ3 system. 2.2 Speech Recognition Speech recognizers are computer systems that process human speech into something a computer can recognize and act on. There are several advantages in using speech in an application. For example, you can enter data when no keyboard or screen is available. It is also very convenient to use speech when hands or eyes, or both, are busy, and in difficult environments such as darkness or cold. Some areas where speech recognition is useful are: help for functional disabled people who are not able to type using their hands, telephone services where you have a very limited keyboard and when you need free hands, for example talking in your cell phone while driving speech recognizers into different categories. First divide them into speaker dependent or speaker independent. Natural-language speech recognition refers to computer systems that recognize and act on unconstrained speech. That is, the user does not need to know a predefined set of command words in order to use the system successfully (Boyce, 2000). Good speech recognition can be quite hard to achieve. This makes it difficult to find the word boundaries, compared to finding them in written text, which makes the words difficult to recognize. There is a great variability in speech between speakers depending on age, sex and dialect and speech within a speaker depending on mood, stress etc. External conditions such as background and recording device also make a difference. 2.3 Speech Synthesis Speech synthesis is when the computer transforms textual or phonetic description into speech. A Text-To- Speech (TTS) synthesizer is a computer-based system that should be able to read any text aloud. There is a fundamental difference between this kind of system and any other talking machine (a cassette-player for example) because of the automatic production of new sentences that a TTS can perform. 2.4 Existing Voice-Controled Systems Most car companies do some kind of research about voice-controlled systems today. BMW for example has voice-control in most of its vehicles7. Volvo is another car company that works with voice-controlled solutions in the car. 2.5 Syntactic Analysis We did these transcriptions in order to be able to classify everything our test persons said into the different functions they wanted performed with their utterances, or the functions their utterances were related to. An utterance is a string of speech found between breath and pauses. Every test person s utterances were divided into stereoand address book functions. The utterances which were related to the stereo, were divided into these categories: Change the volume Change tune/cd Turn on/turn off The utterances which were related to the address book, were divided into these categories: Missed calls Check address book

293 271 Add to address book Delete Change We also divided the utterances to the computer system and the utterances to the human system into different columns. We were mostly interested in the utterances made to the computer system, since this is what we believe to be what a voice controlled system has to handle. This, we believe, is also very likely to be the way people will interact with a voice-controlled computer system. We were also interested in comparing the utterances to the computer system with the utterances to the human system. As can be seen from the comments to Table No.2 the utterances from the computer system and the test leader are very similar. Table 1: An example of utterances from a test person divided into respective functions 7 FV AdvP ({NP, NP PP}) 5 FV NP PP (NP) 4 AdvP NFV NP FV (AdvP) PP 4 AdvP FV NP ({AdvP, NP AdvP}) 3 AdvP NFV NP NFV FV {AdvP, PP, AdvP NP} 3 NP FV {AdvP, NP, AdvP NP } 2 NFV NP FV (AdvP) 2 InterjP 2 NP NFV NFV FV {NP PP, PP} 2 FV NP (NP) PP NP NP NFV NP 1 AdvP NFV NP FV NP NP NFV PP 1 AdvP FV NP AdvP NP PP NP NP NP NFV NP 1 AdvP FV NP PP 1 AdvP FV PP NP NP NFV NP PP 1 NFV FV AdvP NP PP 1 NFV FV NP 1 NP NFV AdvP NFV FV PP 1 NP NFV NFV FV AdvP NP NP 1 NP NFV FV PP 1 FV NP PP NP NP PP NP NP 1 FV PP NP NP NFV NP 56 Fig. 1 The individual difference in words/utterance between the two tests. 3. Speech Recognition 2.6 Syntax of sentences uttered to computer system 46 NP ({PP, NP, FV NP}) 39 FV PP ({NP, PP, AdvP, PP NP}) 36 FV NP ({NP, AP, PP, AdvP, NP NP, NP PP}) FV (AdvP) NP Speech recognition has a history of more than 50 years. With the emerging of powerful Computers and advanced algorithms, speech recognition has undergone a great amount of progress over the last 25 years. The earliest attempts to build systems for automatic Speech recognition (ASR) was made in 1950s based on acoustic

294 272 phonetics. These systems relied on spectral measurements, using spectrum analysis and pattern matching to make Recognition decisions, on tasks such as vowel recognition [1]. Filter bank analysis was also utilized in some systems to provide spectral information. In the 1960s, several basic ideas in speech recognition emerged. Zero-crossing analysis and speech segmentation were used, and dynamic time aligning and tracking ideas were proposed [2]. In the 1970s, speech recognition research achieved major milestones. Tasks such as isolated word recognition became possible using Dynamic Time Warping (DTW). Linear Predictive Coding (LPC) was extended from speech coding to speech recognition systems based on LPC spectral parameters. IBM initiated the effort of large vocabulary speech recognition in the 70s [3], which turned out to be highly successful and had a great impact in speech recognition research. Also, AT&T Bell Labs began making truly speaker-independent speech recognition systems by studying clustering algorithms for creating speaker-independent patterns [4]. In the 1980s, connected word recognition systems were devised based on algorithms that concatenated isolated words for recognition. The most important direction was a transition of approaches from template-based to statistical modeling especially the Hidden Markov Model (HMM) approach [5]. HMMs were not widely used in speech application until the mid-1980s. From then on, almost all speech research has involved using the HMM technique. In the late 1980s, neural networks were also introduced to problems in speech recognition as a signal classification technique. Recent focus is on large vocabulary, continuous speech recognition systems. 3.1 Acoustic Model by Hidden Markov Model Any acoustic unit, such as word, syllable, diphone, triphones or phone can be modeled by Hidden Markov Models (HMMs). An HMM is a finite state machine which can be viewed as a generator of random observation sequences according to probability density functions. The model changes state once at each time step and at time t a state j is entered, a speech vector o t is generated from the probability density a ij. The values of aij should satisfy N j 1 a 1 (1) ij where N is the number of states. They provide the temporal information in the HMM [6]. The quantity P(Y/W) is the probability of an acoustic vector sequence Y given a word sequence W to find the most probable word sequence. A simplistic approach to achieve this would be to obtain several samples of each possible word sequence, convert each sample to the corresponding acoustic vector sequence and compute a statistical similarity metric for the given acoustic vector sequence Y to the set of known samples. For large vocabulary speech recognition this is not feasible because the set of possible word sequences is very large. Instead words may be represented as sequences of basic sounds. Knowing the statistical correspondence between the basic sounds and acoustic vectors, the required probability can be computed. Fig. 3 Trip hones HMM 3.2 Deccoder or Recognizer Fig. 2 Messages Encoding and Decoding in a ASR A decoder is a searching algorithm, which finds the corresponding word sequence. We given the maximum a posteriori probability P (W O) for a spoken utterance O and its corresponding word string W. In the HMM based

295 273 recognition system, decoding is controlled by a recognition network. A recognition network consists of a word-level network, a dictionary and a set of HMMs. A word network describes the sequence of words that can be recognized and, for the case of sub-word systems, a dictionary describes the sequence of HMMs that constitute each word. A word-level network will typically represent either a finite-state Task Grammar which defines all of the legal word sequences explicitly or a Word Loop which simply puts all words of the vocabulary in a loop and therefore allows any word to follow any other word. Word-loop networks are often augmented by a stochastic language model (LM). A recognition network ultimately consists of HMM states connected by transitions. Fig. 5 Steps Sequence in Converting a speech signal into a set parameters suitable for ASR 3.3 Parameterization Fig. 4 Recognition Network Model For input into a digital computer the continuous speech signal must first be converted into discrete samples which are then converted into a set of representative feature vectors. This parameterization process is often called the front-end of the speech recognition system. The steps involve in converting speech signal into a set of parameters are shown in Figure 5. The main purpose of the digitization process is to produce a sampled data representation of the speech signal with as high a Signal to Noise ratio (SNR) as possible [11]. The process of grouping digitalized speech into a set of samples, called frame, typically represented between 20 and 30 ms of speech. The digitized speech signal is blocked into overlapping frames [1]-[20] as shown in Fig. 10. The overlap decreases problems that might otherwise occur due to signal data discontinuity. A one coefficient digital filter, known as a Preemphasis filter[11]. This stage spectrally flattens the frame using a first order filter. The transformation may be described as[12]: Here, X[n] refers to the speech sample in the frame. Sphinx uses α= 0.97 and the sampling rate is typically 8K or 16K 16-bit samples per second. Windows are functions defined across the time record which are periodic in the time record. They start and stop at zero and are smooth functions in between. When the time record is windowed, its points are multiplied by the window function, time bin by time bin, and the resulting time record is by definition periodic. It may not be identical from record to record, but it will be periodic (zero at each end). In the frequency domain, a window acts like a filter. The amplitude of each frequency bin is determined by centering this filter on each bin and measuring how much of the signal falls within the filter. If the filter is narrow, then only frequencies near the bin will contribute to the bin. A narrow filter is called a selective window, it selects a

296 274 small range of frequencies around each bin. However, since the filter is narrow, it falls off from center rapidly. This means that even frequencies close to the bin may be attenuated somewhat. If the filter is wide, then frequencies far from the bin will contribute to the bin amplitude but those close by will probably not be attenuated much. The net result of windowing is to reduce the amount of smearing in the spectrum from signals not exactly periodic with the time record. The different types of windows trade off selectivity, amplitude accuracy, and noise floor. Today, in speech recognition, the Hamming window is almost exclusively used. The Hamming window is a specific case of the Hanning window [11]. In this stage a Hamming window is applied to the frame to minimize the effect of discontinuities at the edges of the frame during FFT. The transformation is[12]: logic 1 is electrically represented by 5 V and logic 0 is represented by 0 V (digital ground). Whenever the Data and Clock line is not used, i.e. is idle, both the Data and Clock lines are left floating, that is the host and the device both set the outputs in high impedance. Externally, at the PCB, large (about 5 k ) pull-up resistors keep the idle lines at 5V (logic 1). The FPGA/keyboard interface is shown in figure 7. When the FPGA reads the Data or Clock inputs both PS2Data_out and PS2Clk_out are kept low which puts the tri-state buffers in high impedance mode. When the FPGA "writes" a logic 0 on an output, the corresponding x_out (x = PS2Data or PS2Clk) signal is set high which pulls the line low. When writing logic 1 the FPGA simply sets the x_out signal low. The vector H[n] is computed using the following equation[12]. The constants used in the H[n] transform were obtained from the Sphinx source code. In practice, it is desirable to normalize the window so that the power in the signal after windowing is approximately equal to the power of the signal before windowing. The purpose of the window is to weight, or favor, samples towards the center of the window [11]. Speech coding is the compression of speech into a code using audio signal processing and speech processing techniques. To encode the speech signal into a suitable set of parameters three basic classes of techniques are being used: Fig. 7 FPGA/ Keyboard Interface 4.1 Protocol for receiving data from the keyboard Data is received from the keyboard as illustrated in Figure 8. Fourier transformations Filtering through digital filter-banks Linear prediction Fig. 8 PS/2 protocol Since the speech signal is not stationary, speech analysis for encoding must be performed on short-term windowed segments, usually, a duration of 20 to 30 ms with a frame period of 10 to 15 ms. Therefore, for short period of time (~10ms) the speech signal is quasistationary and this allows us to represent the signal over this period by a single feature vector. 4. Keyboard Interface The PS/2 interface is a bit serial interface with two signals Data and Clock. Both signals are bi-directional and Fig. 9 PS/2 Timing Data is sent in bit serially. The first bit is always a start bit, logic 0. Then 8 bits are sent with the least significant bit first. The data is padded with a parity bit (odd parity). The parity bit is set if there is an even

297 275 number of 1's in the data bits and reset (logic 0) if there is an odd number of 1's in the data bits. The number of 1's in the data bits plus the parity bit always add up to an odd number (odd parity.) This is used for error detection. A stop bit (logic 1) indicates the end of the data stream. 4.2 The keyboard scan-codes The keyboard sends packets of data, scan codes, to the host indicating which key has been pressed. When a key is pressed or held down a make code is transmitted. When a key is released a break code is transmitted. Every key is assigned a unique make and break code so that the host can determine exactly what has happened. There are three different scan code sets, but all PC keyboards use Scan Code Set 2. A sample of this scan code set is listed in table 2. Please refer to the lab homepage for the full scan code set. 4.3 Displaying scan codes Receive the scan codes from the keyboard and display the corresponding code in hexadecimal format on the XSV board digit LEDs. The LEDs should be updated at a rate of approximately 1 Hz. Following we will partition this problem into more manageable pieces. We will partition the design into a data path and a control path. A block diagram of the complete design is shown in figure 10. Table 2: Scan Code Set 2 (sample) KEY MAKE BREAK A 1C FO, 1C B 32 FO, 32 C 21 FO, 21 D 23 FO, 23 E 24 FO, 24 F 2B FO, 2B G 34 FO, 34 H 33 FO, 33 I 43 FO, 43 J 3B FO, 3B K 42 FO, 42 L 4B FO, 4B M 3A FO, 3A N 31 FO, 31 O 44 FO, 44 P 4D FO, 4D Q 15 FO, 15 R 1B FO, 2D S 1B FO, 1B T 2C FO, 2C U 3C FO, 3C V 2A FO, 2A W 1D FO, 1D X 22 FO, 22 Y 35 FO, 35 Z 1A FO, 1A Fig. 10 A block diagram of a PS/2 Keyboard Interface All flip-flops in the design are clocked with a 20 MHz clock and the rising edge is the active edge. Four types of positive edge-triggered D-flip-flops are shown in figure 4.FD No reset/set FDC Asynchronous reset FDP Asynchronous set FDCE Clock enable and asynchronous reset The PS2Data and PS2Clk signals are sampled with FDP flip-flops. The PS2Data is fed into a shift register. When 8 data bits (scan code) and one parity bit has been shifted into the shift register the scan code is written to a synchronous FIFO. When the FIFO is not empty the output is read in intervals of approximately 1 second. The FIFO output is connected to two identical ROMs. The ROMs decode the scan code data so that a character 0 F is displayed on each digit LED.A short description of

298 276 each module in the data and control path follows. The keyboard scanning process is illustrated in Figure 8. First the column result is scanned. The three lower bits (column part) of the scan code is incremented until the low column line is found. Then, the port directions are inverted and the row result is scanned. The row part of the scan code is incremented until the low row bit is found. Finally, the scan code processing function is called. 5. Speech Application Programming Interface The Speech Application Programming Interface (SAPI) is defined by Microsoft. Version 4 was published in the beginning of 1998 and its successor, version 5, was published in the beginning of Maybe the biggest weakness in SAPI 4 is the fact that there is no centralized control panel for speech synthesizer parameters. This means that user is required to select his or her favorite voice and adjust it individually for every speaking application. SAPI 5 adds the Speech item to the control panel. This control panel item is used by the user to define his or her favourite voice and other speech parameters SAPI (Speech Application Programming Interface) was first introduced to Windows 95. This API provides a unified interface for dynamic speech synthesis and recognition. Over the years new versions were developed and now it is version 4.0 with WinXP. Unfortunately the API wasn't really maturated and supported only C++ (later Visual Basic and COM), so it was quite widely used. Microsoft redesigned the version 5.0 from scratch and changed critical parts of the interface. However the latest stable version 5.1 is still a native code DLL, but with the next one, which is considered to be part of Windows Vista (a complete redesign again), A full support for managed.net code will be expected [7]. Right now it is only possible to take advantage of the current SAPI interface via C# by using COM Interop, which is.net technique to use native COM object Programming interfaces expose the full power of complex software engines. SAPI 5 is Microsoft s fifth iteration towards a speech application programming interface to enable programmers to build speech-enabled applications for Microsoft Windows platforms. It incorporates lessons that Microsoft has learned over many years and many APIs, in an effort to make the best comprehensive API possible. Even so, most programmers should do at least their prototyping and in most cases their product using a tool suite that does most of the SAPI 5 programming for them. Why? As always, to improve value and reduce costs. Microsoft SAPI is a little known speech recognition and synthesis engine that is supported in all versions of Windows after Windows 95. The Microsoft Speech SDK (System Development Kit) version 5.1 is available for free download from the Microsoft Speech Technologies Website. Speech SDK 5.1 is compatible with a number of programming languages, but most of the documentation focuses on examples written in C++. In general, any programming environment that supports OLE automation will work for writing SAPI applications. 6. Result and Discussion In our project work, an attempt has been made to develop voice control computer system. Here computer is the central device. Microphone is the input device and voice is the input. For this system a microphone is connected to the pc via sound cord and then software was developed to accept the input processed by sound application programming interface(sapi) software. Input was processed by microprocessor and display this. We do believe that a voice-controlled system must be able to handle some kinds of disturbances in the utterance. The system must not interpret them as speech or as the ending of an utterance. This is especially important when the system is going to be used in a car since this is an environment where these disturbances can often occur since your attention is on the traffic and everything around you. When something happens that catches your attention you might, for example, make unmotivated pauses in your utterance to the computer, which the computer system does not have a clue why you are doing and probably will interpret it as if you are finished with your utterance. To actually build a voice-controlled system on the basis of our communication model and evaluate it would also be a great challenge. Thereafter, to do a study with a large group of test persons to find out how our system would work, would have been challenging. To do a study with a large group of test persons would also be interesting since we could see if the results could be statistically secure. When we created our communication model we worked on the basis of the specific environment of the car. Therefore our solution is custom-made for the car. However, it might be possible to use it for other areas where you use service gateways. For example when you want to use voice-controlled computers in your home environment. System performance totally depends on the output of the system. The percentage of success rate and failure rate has been calculated using the following equations: Success: Total number of Total number of success Test Rate= 100%

299 277 Failure: Total number of Total number of failure Test Rate= 100% The performance is related to success rate and failure rate. If the success is high then the performance of the system is good. Success rate and Failure rate are contradiction of each other. So when success rate is high then failure rate is low. In the two terms the performance of the system is depended. 7. Conclusions We have demonstrated a voice control system, for physically disabled persons, that operates with a minimum of specialized hardware. Some of the chosen solutions are dictated by the wish to build on the equipment already used by our test subject. For a more mobile person, the same solution could be used in connection with a portable computer. The telephone capability could be given a less expensive and more flexible solution through a standard modem. The use of a standard PC gives potential access to a wide variety of programs. Future expansions include improved software for text processing. Acknowledgments The authors would like to thank student Pulack Chakma [M.Sc.(Tech.)] for her assistance during the computer simulations and Md. Farzan Ali [M.Sc. (Mat.)] for the language checking. References [1] Visual Basic v6 (Part 1-Database and Multimedia Programming) By Mahabubur Rahaman [2] Visual Basic v6 (Part 2- Database and Multimedia Programming) By Mahabubur Rahaman [3] Mastering Visual Basic 6 By Michael j. Young [4] Mastering Visual Basic 6 By Evangelos petroutsos [5] Microprocessor and interfacing programming and hardware By Douglas Hall,Tata McGraw-Hill Edition [6] Microprocessors and Microprocessors based system design By Dr. M.Rafiquzzaman [7] Spectral contrast normalization and other techniques for speech recognition in noise,. C. Bateman, et. Al [8] Speech enhancement based on masking properties of the auditory system [9] H. Hermansky, Perceptual linear predictive (PLP) analysis of speech, The Journal f the Acoustical Society of America, vol. 87, no. 4, pp , 1987 [10] B. Atal and M. Schroeder, Predictive coding of speech signals, Proc. of 6 th International Congress on Acoustics, Tokyo, pp , 1968 [11] H. Matsumoto and M. Moroto, Evaluation of Mel-LPC cepstrum in a large ocabulary continuous speech recognition, Proc. ICASSP, pp , 2001 [12] P. H. Lindsay and D. A. Norman, Human information processing: An introduction to psychology, 2nd Ed., pp. 163, Academic Press, 1977 [13] S. F. Boll, Suppression of acoustic noise in speech using spectral subtraction, IEEE Trans. Acoust., Speech and Signal Processing vol. 27, no. 2, pp [14] H. W. Strube, Linear prediction on a warped frequency scalle, J. Acoust. Soc. Am., vol. 68, no. 4, pp , 1980 [15] H. Matsumoto, Y. Nakatoh and Y. Furuhata, An efficient Mel-LPC analysis method for speech recognition,proc. ICSLP 98, pp , 1998 [16] H. Matsumoto and M. Moroto, Evaluation of Mel-LPC cepstrum in a large vocabulary continuous speech recognition, Proc. ICASSP, pp , 2001 [17] Sub Auditory Speech Recognition" by Kim Binsted and Charles Jorgensen [18] "Sub Auditory Speech Recognition Base on EMG/EPG Signals" by Chuck Jorgensen, Diana D. Lee, and Shane Gabon [19] S. Itahashi and S. Yokoyama, A formant extraction method utilizing mel scaleand equal loudness contour, Speech Transmission Lab.-Quarterly Progress and Status Report (Stockholm) (4), pp , 1987 [20] Various websites On Internet Md. Sipon Miah received the Bachelor s and Master s degree in Information and Communicatio Engineering from Islamic University, Kushtia, 2006 and 2007, respectively. He is currently Lecturer in the department of Information and Communication Engineering, Islamic University, Kushtia,Bangladesh.Science 2003, he has been a Research Scientist at the Communication Research Laboratory, the department of ICE, Islamic University, Kushtia, where he belongs to the spread-spectrum research group.he is pursing research in the area of internetworking in wireless communication. He has three published papers in international and national journals in the same area but some papers under process. His areas of include database system, interfacing, programming, optical fiber communication and wireless communications. Tapan Kumar Godder received the Bachelor s, Master s and MPh degree in Applied Physics & Electronics from Rajshahi University, Rajshahi. In 1994, 1995 and 2007, respectively. He is courrently Associate Professor in the department of ICE, Islamic University, Kushtia-7003, Bangladesh. He has seventeen published papers in international and national journals. His areas of interest include internetworking, AI & mobile communication.

300 278 Combining of Spatial and Frequency Domain Transformation With The Effect of Using and Non-Using Adaptive Quantization for Image Compression Alan Anwer Abdulla Mathematics Dept, University of Sulaimani, Sulaimani, IRAQ Abstract Usage of image has been increasing and used in many applications. Image compression plays vital role in saving storage space and saving time while sending images over network. One important point of delivering digital image is time. In this paper we concentrated on the speed of packet sending with reasonable quality. The aim of this paper is to reduce image size with reasonable quality. By getting advantage from image transformation in both scope spatial domain and frequency domain. Splitting the image into 4 blocks in the case of spatial domain and applying Standard Haar wavelet transformation in the case of frequency domain was the main factor for approaching the aim of this paper which was reduce image size and increase speed of delivering digital image. After huge tests in the case of frequency domain in all subbands(ll,hl,lh,hh) we made a decision for which subband contains high information and which subband contains poor information. Depending of these factors we could send less than oneeighth of the size of the original image to the destination and getting the reconstructed image with the acceptable quality as a result of this paper. Keywords: -Haar Wavelet Transformation; Adaptive Quantization; Spatial Domain; Frequency Domain; PSNR ; Mean Square Error; Compression Ratio; Variable Encoding RLE; Image Quality 1. Introduction The image compression highly used in all applications like medical imaging, satellite imaging, etc. The image compression helps to reduce the size of the image, so that the compressed image could be sent over the computer network from one place to another in short amount of time. Also, the compressed image helps to store more number of images on the storage device [1,2]. Image compression plays a critical role in telemetric applications. It is desired that either single image or sequences of images be transmitted over computer networks at large distances so that they could be used for a multitude of purposes. For instance, it is necessary that medical images can be transmitted so as to be reliable, improved and fast medical diagnosis performed by many centers could be facilitated. To this end, image compression is an important research issue [3]. Images contain large amounts of information that requires much storage space, large transmission bandwidths and long transmission times. Therefore it is advantageous to compress the image by storing only the essential information needed to reconstruct the image. An image can be thought of as a matrix of pixel (or intensity) values. In order to compress the image, redundancies must be exploited, for example, areas where there is little or no change between pixel values. Therefore images having large areas of uniform color will have large redundancies, and conversely images that have frequent and large changes in color will be less redundant and harder to compress. Wavelet analysis can be used to divide the information of an image into approximation and detail subsignals. The approximation subsignal shows the general trend of pixel values, and three detail subsignals show the vertical, horizontal and diagonal details in the image. The aims behind the adaptive quantization are to quantize the retained coefficients after transformation step according to the quantity of information existed in each subbands and to obtain a large sequence of zeros especially in (HL, LH and HH) bands [4]. In general, Figure 1 shows all steps of the work in this paper as a diagram. Steps are including: 1- Convert the original image in the case of spatial domain by taking average mean of neighboring and getting (original image2). 2- Splitting the (original image2) into 4 blocks in the case of spatial domain. 3- Putting zero value for all blocks except the top-left block which remain as its. 4- Applying Standard Haar wavelet (2 nd level) on the remained block (Top-left). 5- Putting zero value for eight blocks out of 16 blocks. 6- Applying adaptive quantization on 4 blocks out of 8 remained blocks. 7- Apply variable run length encoding to find compression ratio.

301 279 Raw Image 2. Methods and Materials Procedure 2.1. Spatial Domain Conversion Convert Original Image by taking average mean of neighboring and getting (original image2) Splitting the (original image2) into 4blocks Putting zero value for all blocks except the Top - left block Applying Standard Haar wavelet (Encoding) at (2nd level) on the remained block (Top-left) Putting zero value for eight blocks out of 16 blocks The term spatial domain refers to the image plane itself, and approaches in this domain are based on direct manipulation of pixels in an image. The spatial domain is aggregate of pixels composing an image. Spatial domain methods are procedures that operate directly on these pixels. Spatial domain processes will be denoted by the expression g( x, y) T f ( x, y) Where f ( x, y) is the input image, g( x, y) is the processed image, and T is an operation on f,defined over some neighborhood of ( x, y) [5]. In this paper the spatial domain conversion done in two steps which are discussed in , and Average Mean of Neighboring It is one of the methods that operate on the pixels directly. By taking the average value of 4 neighbor pixels and put this average value for each one of these 4 neighbor pixels. After applying average mean on original image (shown in Figure.2) the image result shown in Figure.3. Applying adaptive quantization on HL and LH subbands only Applying Standard Haar wavelet (Decoding) De-splitting the (original image2) from 4block into one block (image) Reconstructed Image Figure 1. Diagram of paper s procedure Figure 2. Original image

302 280 Figure 3. Image after applying average mean Splitting image(figure 3) into 4 Blocks It is another method that operates on the spatial domain. In this step, image splits into 4 blocks with equal sizes (shown in Figure.4). The aim of splitting is to work only on one block and ignore other three blocks in order to reduce size. Figure 5. Putting zero to three blocks 2.2 Frequency Domain Conversion Frequency domain conversion includes: Applying Standard Haar wavelet (Encoding) By applying standard Haar wavelet at 2nd level on the remained block (Top-left block of Figure.5), the image transfer from its spatial domain to frequency domain. At the 1st level the frequency domain includes 4 subbands (LL,HL,LH,HH) then by applying wavelet again on each subband, the frequency domain includes 16 subbands (LL of LL1, HL-LL1, LH-LL1, HH-LL1,LL-HL1,HL-HL1,LH- HL1,HH-HL1,LL-LH1,HL-LH1,LH-LH1,HH-LH1,LL-HH1,HL- HH1,LH-HH1,HH-HH1) as shown in Figure 6. Figure 4. Splitting image(in Figure.3) into 4 blocks Putting zero value By putting zero values to three blocks except top-left block as third step in the spatial domain (shown in Figure.5) as ignoring these three blocks. Figure 6. After applying Standard Haar Wavelet (2 nd level)

303 Putting zero to the subbands By putting zero to the subbands that contains poor information. After a lot of tests, show that eight subbands contain poor information. For example HL subband contains vertical information,then by applying wavelet again the LH of HL contains poor information because there is no horizontal information in the HL subband. According to these tests, we put zero value to eight subbans out of 16 subbands as shown in Figure 7. ( ) parameter was varied between the values of alpha 01). Adaptive algorithm for forward quantization is as follows: Q _ f round Y ( x, y ) ( ) Q _ LL Figure 7. putting zero values to 8 subbands Figure 8. Applying adaptive quantization on,(llof HL,HL of HL,LL of LH,LH of LH) subbands with quantization value 20 for each subband and = Applying Adaptive Quantization Normally, Wavelet data is erratic and very small and large numbers are obtained from applying the DWT. However, much of this data can be discarded with minimal loss to the final image. We apply Quantization to wavelet data to achieve this. Quantization is a very important part of compression when dealing with wavelets [6]. In these tests, a universal scalar step size is used for quantization of all the packets. The amount of compression went up significantly when more quantization was applied. This is due to the greater amount of zeros after quantization (zeros are encoded very efficiently). In this paper, adaptive quantization applied on HL and LH subbands only in the previous image, because all subbands of LL in the 2nd level (LL of LL, HL of LL, LH of LL, HH of LL) contain important information, by applying adaptive quantization the most important information will be lost. As you see, in the previous image all subbands of HH (LL of HH,HL of HH,LH of HH, HH of HH) ignored (set zero for all subbands), so it doesn t need to apply quantization on it. The remained subbands that need to apply quantization are four subbands out of 8 subbands which are,(ll of HL,HL of HL,LLof LH,LH of LH) as shown in Figure.8 with quantization value 20 and = 0.5 (while Applying Standard Haar wavelet (Decoding) After applying standard Haar wavelet decoding on previous image(in Figure 8), the reconstructed image shown in Figure 9. Figure 9. Reconstructed image

304 282 Then by de-converting the reconstructed image (in Figure 9) from one block to 4 blocks the image becomes: 3. Experimental study and discussion of the results Figure 10. De-converting from one block to 4 blocks Then by de-splitting the image(in Figure 10) from 4 blocks into one image. The final reconstructed image is: The major effort in this work has been set to the direction of defining the significant and non significant of image compression. The proposed processing scheme tested on several different images (as shown in table 1) but in this paper displayed only one image as an instance by two steps (as shown in Figure.12,13). The first without quantization and the second with quantization. Depending on the image quality measurements (PSNR, MSE) for the first one (without quantization) PSNR and MSE are reasonable which means that the image quality is reasonable, that means by sending 16 packets (subbands) out of 64 packets to the destination, the image quality is very reasonable. For the second step, (with quantization =20, alpha=0.5) the PSNR and MSE are acceptable, that means the image quality is acceptable, that means by sending less than 16 packets out of 64 to the destination, the image quality is acceptable. After that, according to the compression ratio, which supports the previous discussion, we can see the effect of quantization on the speed of delivery and image quality. The compression ratio (C.R.) is the amount of original data divided by the amount of compressed data. The least amount of data needed to represent all of the image information is constrained by the amount of information contained in the values encoded. C. R. amount of original data amount of compressed data The amount of information contained in the values or symbols encoded can be quantitatively measured [7]. Figure 11. Final reconstructed image (one-sixteenth of the original image size) (16 subbands out of 64 subbands)

305 283 Original image Original image 16 subbands out of subbands out of 64 with quantization= 20, alpha= 0. 5 Reconstructed image PSNR= 29 MSE= 9 C.R=1.95 Figure 12. without quantization Reconstructed image PSNR= 24 MSE= 31 C.R= 2.77 Figure 13. with quantization

306 284 Original image Table 1: RESULTS OF C.R, MSE AND PSNR BETWEEN ORIGINAL IMAGE AND RECONSTRUCTED IMAGE WITHOUT AND WITH QUANTIZATION Without quantization With quantization MSE PSNR C.R MSE PSNR C.R [6] Khalid Sayood, Introduction to Data Compression, Morgan Kaufmann Publishers, San Francisco California, [7] Steven C. Meadows, Color Image Compression Using Wavelet Transformation, Texas Tech University, Image1 Image2 Image3 Image4 Image Conclusions The present experiment reveals that the proposed procedure achieves better compression ratio with using adaptive quantization than without using. The tested results show that the quality of reconstructed image is better in the case of without using adaptive quantization which was PSNR= 29 db. This is contrast of the case of using adaptive quantization which was PSNR = 24 db, but the C.R was better. In this paper, Standard Haar wavelet used for the second level because in Standard all subbands are equal in size, as we treat each subband as a packet so the Standard wavelet was useful for this issue. References [1] Gonzalez, R.C. and Woods, R.E., Digital Image Processing 2nd ed., Pearson Education, India, [2] Singara Singh, R. K. Sharma, M.K. Sharma, Use of Wavelet Transform Extension for Graphics Image Compression using JPEG2000 Framework, International Journal of Image Processing, Volume 3, Issue 1, Pages 55-60, [3] J. Meunier and M. Bertrand, «Ultrasonic texture motion analysis: Theory and Simulation,» IEEE Trans. Med. Imaging, vol. 14, no. 2, pp , June [4] K. Ramkishor et al, Adaptation of Video Encoders for Improvement in Quality, IEEE International Symposium in Circuits and System, Vol.2, 2003, pp [5] Jonathan M. Blackledge, Digital Image Processing Mathematical and Computational Methods.Loughborough University, England, 2005.

307 285 Automatic Computation for Pressure Controlled Intermittent Coronary Sinus Occlusion Loay Alzubaidi 1, Werner Mohl 2, Frank Rattay 3 1 Department of computer science, Prince Muhammad bin Fahd University AL-Khobar, Saudi Arabia 2 Department of Cardiothoracic Surgery, Medical University of Vienna Vienna, Austria 3 Institute for Analysis and Scientific Computing, Vienna University of Technology Vienna, Austria Abstract Pressure controlled intermittent coronary sinus occlusion (PICSO) has been found to substantially salvage ischemic myocardium. By elevating venous pressure two mechanisms are involved namely the distention of venous vessels inducing mechanotransduction as well as a redistribution of venous flow towards ischemic areas. Mechanotransduction in endothelial cells inducing changes in the ischemic heart inducing myocardial salvage as well as myocardial recovery is blunted by possible consequences of myocardial perfusion deficits by limiting coronary inflow. To limit these severe side effects we have evaluated a new mathematical model to describe the Increase (Inflation) and decrease (deflation) in coronary Sinus pressure (CSP) following pressure controlled intermittent coronary sinus occlusion (PICSO) and release. The model is evaluated and compared on the basis of dogs, pigs and sheep. The model consists of two parts with three parameter double exponential function for each, and it was fitted by using the non-linear least squares algorithms. The new model was used in implementation of automatic computing module which is responsible to compute the following quantities for Inflation and Deflation: 1. Systolic and diastolic plateau. 2. Rise-Time of systolic and diastolic plateau. 3. The mean integral of the CSP (Area under the curve). 4. Number of heart beats impact in inflation and deflation periods. 5. Driving the slope of CSP Corresponding quantities for various coronary sinus balloon inflation and deflation cycles were computed with ranging from cycles being as short as 5sec/3sec (inflation/deflation ratio) to cycles as long as 12sec/8sec. Keywords: coronary sinus pressure (CSP), pressure controlled intermittent coronary sinus occlusion (PICSO) and Left Anterior Descending Artery (LAD). 1. Introduction Pressure controlled intermittent coronary Sinus occlusion PICSO has been proposed and investigated as a new technique in interventional cardiology and cardiac surgery to salvage ischemic areas of the myocardium. CSP in mmhg inflation deflation time in sec Fig. 1 One PICSO with Inflation and deflation time 10/6 sec Because the beneficial effect of this intervention appears to be closely linked to an optimal timing of coronary venous pressure elevation (i.e. to define the therapeutic range of salvage and severe side effects induced by prolonged increase in coronary resistance), a new mathematical model has been developed in order to put the estimation of occlusion and release times on a quantitative basis. The model consists of two parts with three parameter double exponential function for each.

308 286 A* exp {B*[ 1 exp ( C*t)] 1} when 0 t T1 Pcsp(t) F. ( 1) D* exp {E*[ 1 exp ( )] 1} when T1 t T 2 t Where Pcsp (t) = Coronary sinus pressure (mmhg) t = Time (s), measured from the start of occlusion A, D =fitting parameter in (mmhg) B, E = fitting parameter (dimensionless) C, F= Fitting parameter (1/s) T1 = Time that mark the end of the CSP occlusion phase (s) T2 = Time that mark the end of the CSP release phase (s) The first part Eq. (1a) describes the rise of the CSP during the Inflation (occlusion) time. Pcsp (t) A* exp {B*[ 1 exp ( C*t)] 1} ( 1a) As shown in the Fig. (1), systolic peaks incremented coincided with the time during the inflation period. The second part Eq. (1b) describes the release of the CSP during the deflation (release) time. F Pcsp (t) D* exp {E*[ 1 exp ( )] 1 } ( 1b) t As shown in the Fig. (1), the systolic peaks decremented coincided with the time during the deflation period. The systolic and diastolic peaks were fitted with the nonlinear least least-square algorithms. Additionally a new module was implemented using Oracle DB and Matlab, which is responsible to compute automatically the quantities of the CSP. 2. Method Since the mathematical model Eq. (1a and 1b) together with both sets of three fitted parameters (A,B,C and D,E,F) represent the envelope curve (systolic and diastolic, respectively), it is now possible to express the height of the CSP plateau as well as the time taken to reach the plateau in terms of fitted parameters. 2.1 Inflation The inflation haemodynamic quantities were calculated by using the Eq. (1a) which developed by Schreiner [W. Schreiner] CSP plateau and rise time The highest value of Pcsp (t) in Eq. (1a) is reached for t Pcsp ( t ) A *exp (B -1) Since a plateau is never actually reached (in mathematical terms) it is meaningful to consider the time it takes to reach, say, 90% of the predicted height of the plateau. The systolic Plateau is 90% from the Pcsp ( t ) Systolic Plateau 0. 9 * A * exp (B-1 ) ( 2 ) The Diastolic Plateau was calculated exactly as the Systolic. The rise time will be Rise Time ( 1/C) * ln (-B/ ln ( 0. 9 ) ) ( 3 ) The diastolic plateau and rise time was calculated exactly as the systolic plateau. CSP in mmhg Systolic Plateau Time in Sec Fig. 2 Systolic Plateau of CSP during the inflation, the solid curve represents the fitted three parameter model functions, and the circle represents the systolic plateau Driving the slope of the CSP The first derivative gives an estimate for the maximal slope in terms of the fitted parameters dp/dt A*B*C* exp(b*( 1- exp(-c*t))-c*t) ( 4 ) 2.2 Deflation The deflation was calculated by using Eq. (1b) CSP plateau and rise time The lowest value of Pcsp (t) in Eq. (1b) is reached for t Pcsp(t ) D * exp(-1 ) Since a plateau is never actually reached (in mathematical terms) it is meaningful to consider the systolic Plateau is 110% of the predicted lowest of the plateau. The systolic plateau is 110% from the Pcsp ( t ) Systolic Plateau 11. * D * exp(-1 ) ( 5 ) The rise time will be Rise Time -F / ln ( 1 - ln ( 11. ) / E) ( 6 )

309 287 The diastolic plateau and rise time was calculated exactly as the systolic plateau systolic and diastolic plateau for inflation Fitting Curve 50 CSP in mmhg CSP in mmhg systolic plateau for inflation Systolic Plateau Time in sec Fig. 3 Systolic Plateau of CSP during the deflation, the solid curve represents the fitted three parameter model functions, and the circle represents the systolic plateau Driving the slope of the CSP The first derivative gives an estimate for the maximal slope in terms of the fitted parameters dp/dt ( 1/t ² ) D*E*F* exp( 1- exp(-f/t))*e-f/t -1 ( 7 ) 3. Results The module computes automatically the haemodynamic quantities of the CSP in Multi-cycle, first finding the systolic and peaks for deflation and inflation then making the fitting curve and finding the plateau, rise-time, slope of CSP, mean integral, and the number of heart beats. The model parameters and the derived quantities will change with time. For any diagnostic value it is essential to establish ranges which can be used as reference intervals for the normal state. The following results comprise a preliminary investigation of the spread of the derived quantities observed during PICSO. 3.1 Systolic and diastolic plateau The Automatic computation module computes the systolic and diastolic plateau of the CSP for inflation and deflation periods. The plateau of coronary sinus pressure during LAD occlusion and after LAD reopening for thirteen animals (sheep, dogs and pigs) was computed with this module. Fig. (4) shows the change of the systolic and diastolic plateau of the CSP during the inflation period diastolic dlateau for inflation Time in sec Fig. 4 Systolic and diastolic plateau during the inflation period Statistical result of changing the systolic plateau during the inflation period depending on the occluded or opened LAD is illustrated in Fig. (5). CSP in mmhg Systolic plateau of csp during the inflation period reperfusion early reperfusion later occluded early occluded later Fig. 5 Systolic pressure changes in dependence of LAD during the inflation period The systolic plateau of the CSP will reached ± 5.00 mmhg by reperfusion early (LAD opened), and ± 3.00 mmhg by reperfusion late (LAD opened) and by occluded early (LAD occluded) and ± 2.50 mmhg by occluded late (LAD occluded). 3.2 Automatic computation of rise and release time According to the definition, the rise/release time is the time it takes for the CSP to reach its plateau (systolic or diastolic), which can be reached after a prolonged occlusion or release. Fig. 6 demonstrates the different rise times needed to reach the systolic and diastolic plateau in accordance with the LAD occlusion status. During the inflation period the

310 288 systolic plateau can be achieved very fast when LAD is respond. Rise Time in Sec Time in Sec Fig. 6 Rise time for systolic and diastolic plateau of the CSP during the inflation period. Different rise times during the inflation period depending on the occluded or opened LAD for systolic plateau is illustrated in Fig. (7). The systolic plateau will reach in 5.13 ± 1.00 sec by reperfusion early and in 5.31 ± 1.00 sec by reperfusion late and in 5.89 ± 1.20 sec by occluded early and in 5.80 ± 1.10 sec by occluded late. One/twoway ANOVA was to prove the significance of the obvious difference between the LAD status concerning rise times confirmed that the significance is evident (p<0.0083). time in sec Rise time for systolic plateau reperfusion early reperfusion later occluded early occluded later Fig. 7 Rise time of systolic plateau changes in dependence of LAD during the inflation period 3.3 Relations between derived quantities Predicted plateau, rise times, mean Integral of CSP, number of heart beat per PICSO cycle and CSP slope (dp/dt) are calculated from the fitting parameters and may be called the derived quantities. Whereas the parameters themselves are not directly accessible to physiological interpretation, the derived quantities are deliberately constructed to resemble intuitive criteria. Table 1 shows example of the hemodynamic quantities results from the automatic computation module during 3 minutes. Table 1: Hemodynamic quantities result of the automatic computation module Time in sec Systolic plateau in mmhg Rise time in sec Mean integra in mmhg*sec Heart rate beat / PICSO Conclusion The principal goal of this work was to build a robust mathematical model that could accurately and reliably calculate the rise and release of coronary sinus pressure (CSP) during inflation (rise) and deflation (release) periods. Such a model should be a useful tool that could substitute for time-consuming visual inspection of CSP data during heart operations or other surgeries, when time constraints can affect patient outcomes. Currently, during heart surgeries, clinicians must constantly recalibrate CSP data using this visual inspection, making calculations cumbersome and often inaccurate. Further-more, physiological reactions in heart patients change among individuals, and even within the same individual under different conditions. Thus, any reliable model must be able to accommodate such changes and to operate PICSO under optimal conditions. To date, no mathematical model has been developed to predict or describe the relationship of inflation and deflation of the CSP. Schreiner (86) developed a model, which describes the rise of CSP, but without a model that also accounts for the release of the CSP. Without the other half of the model, calculating the CSP has been a matter of guesswork. To solve this problem, we developed a new mathematical model that efficiently describes the rise and the release of the CSP. Under normal

311 289 experimental conditions, one will always find common features that can be exploited. The CSP plateau is such a feature: while the height, the time to reach the plateau, and the maximum slope may vary, nevertheless a plateau is always reached, and this allowed us to identify features that could be incorporated into the model to make it both feasible for necessary physiologic adaptations and stable and robust for calculations. Our model had to meet the following requirements: i. It had to be derived from the minimum number of fitting parameters. ii. It had to estimate the occlusion and release times on a quantitative basis. The model consists of two equations, each having three parameters of double exponential functions. The CSP was expressed in terms of these fitted parameters, three for inflation and three for deflation. The systolic rise time can be used as a calculated parameter for the closed loop regulation of PICSO. The automatic computing module (computer module) calculates and computes several hemodynamic quantities, such as systolic and diastolic plateau, rise time, heart rate per cycle and the mean integral of CSP. This work has a number of implications for bioinformatics. It provides a new mathematical model that describes the increment and decrement of the CSP during individual PISCO cycles. [8] Mohl W, Glogar D, Kenner T, Klepetko W, Moritz A, Moser M. Enhancement of washout induced by pressure controlled intermittent coronary sinus occlusion (PICSO) in the canine and human heart. In: The Coronary Sinus, Vol. 1 Mohl W, Glogar D, Wolner E (editors). Darmstadt: Steinkopf; 1984; pp First Author: Loay Alzubaidi, PhD in computer science from Vienna University of Technology-2004, Department of computer science, Prince Muhammad bin Fahd University, member of the ACM Second Author: Werner Mohl MD PhD, Medical University of Vienna, Department of Cardiac Surgery Third Author: Frank Ratty Ao. Univ. Prof. Dr.techn. Dr.scient.Med., Institute for Analysis and Scientific Computing, Vienna University of Technology 5. References [1] Mohl W, Gueggi M, Haberzeth K, Losert U, Pachinger O, Schabart A. Effects of intermittent coronary sinus occlusion (ICSO) on tissue parameters after ligation of LAD. Bibliotheca Anatomica 1980; 20: [2] Glogar D, Mohl W, Mayr H, Losert U, Sochor H, Wolner E. Pressure-controlled intermitent coronary sinus occlusion reduces myocardial necrosis (Abstract). Am J Cardiol 1982; 49: [3] Schreiner W, Neumann F, Schuster J, Froehlich KC, Mohl W. Computation of derived diagnostic quantities during intermittent coronary sinus occlusion in dogs. Cardiovasc Res 1988; 22(4): [4] Schreiner W, Mohl W, Neumann F, Schuster J. Model of the haemodynamic reactions to intermittent coronary sinus occlusion. J Biomed Eng 1987; 9(2): [5] Kenner T, Moser M, Mohl W, Tied N. Inflow, outflow and pressures in the coronary microcirculation. In: CSI - A New Approach to Interventional Cardiology. Mohl W, Faxon D, Wolner E (editors). Darmstadt: Steinkopff; 1986; 15. [6] Neumann F, Mohl W, Schreiner W. Coronary sinus pressure and arterial flow during intermittent coronary sinus occlusion. Am J Physiol 1989; 256(3 Pt 2): H [7] Moser M, Mohl W, Gallasch E, Kenner T. Optimization of pressure controlled intermittent coronary sinus occlusion intervals by density measurement. In: The Coronary Sinus, Vol. 1. Mohl W, Glogar D, Wolner E (editors). Darmstadt: Steinkopf; 1984; pp

312 290 Computer Science in Education Irshad Ullah Institute, Computer Science, GHSS Ouch Khyber Pakhtunkhwa, Chinarkot, ISO 2-alpha PK, Pakistan Abstract Computer science or computing science (sometimes abbreviated CS) is the learning of the theoretical foundations of information and computation, and of practical techniques for their execution and application in computer systems. It is often described as the efficient study of algorithmic processes that produce, explain, and transform information. In this work, I use Data Mining algorithms from the field of computer science for the analysis process to prove experimentally and practically that how reliable, efficient and fast are these for the analysis of data in education? A solid mathematical threshold (0 to 1) is set to analyze the data. The obtained results will be tested by applying the approach to the databases and data warehouses of different sizes with different threshold values. The results produce will be of different strength from short to the largest sets of data list. By this, we may get the results for different purposes e.g. making future education plan. Key Words of the abstract Computer Science, Education, Results, Education Plan. 1 Introduction 1.1 Computer Science Computer science or computing science (sometimes abbreviated CS) is the study of the theoretical foundations of information and computation, and of practical techniques for their implementation and application in computer systems.[1][2] 1.2 Data Mining Data Mining is the discovery of hidden information found in databases [14] [19]. Data mining functions include clustering, classification, prediction, and associations. One of the most significant data mining applications is that of mining association rules. Association rules, first introduced in 1993 [17], are used to identify relationships among a set of items in databases. The AIS algorithm is the first published algorithm developed to generate all large itemsets in a transaction database [17]. This algorithm has targeted to discover qualitative rules. This technique is limited to only one item in the consequent. This algorithm makes multiple passes over the entire database. The SETM algorithm is proposed in [13] and motivated by the desire to use SQL to calculate large itemsets [18]. In this algorithm each member of the set large itemsets, Lk, is in the form <TID, itemset> where TID is the unique identifier of a transaction. Similarly, each member of the set of candidate itemsets, Ck, is in the form <TID, itemset>. Similar to [17], the SETM algorithm makes multiple passes over the database. The Apriori algorithm [16] is a great achievement in the history of mining association rules. It is by far the most well-known association rule algorithm. This technique uses the property that any subset of a large itemset must be a large itemset. The Off-line Candidate Determination (OCD) technique is proposed in [10], and is based on the idea that small samples are usually quite good for finding large itemsets. Sampling [9] reduces the number of database scans to one in the best case and two in the worst. A sample which can fit in the main memory is first drawn from the database. The set of large itemsets in the sample is then found from this sample by using a level-wise algorithm such as Apriori. Each association rule mining algorithm assumes that the transactions are stored in some basic structure, usually a flat file or a TID list, whereas actual data stored in transaction databases is not in this form. All approaches are based on first finding the large itemsets. The Apriori algorithm appears to be the center of all the association rule mining algorithms. In this work my focus is on association rule mining technique. I take two algorithms, first the well known Apriori and then our own developed SI [12] algorithm.

313 Association Rule Problem A formal statement of the association rule problem is as follows: Definition: [17] [6] Let I = {i 1, i 2,., i m } be a set of m distinct attributes. Let D be a database, where each record (tuple) T has a unique identifier, and contains a set of items such that T I. An association rule is an implication of the form of X Y, where X, Y I are sets of items called itemsets, and X Y =. Here, X is called antecedent while Y is called consequent; the rule means X Y. Association rules can be classified based on the type of vales, dimensions of data, and levels of abstractions involved in the rule. If a rule concerns associations between the presence or absence of items, it is called Boolean association rule. And the dataset consisting of attributes which can assume only binary (0-absent, 1-present) values is called Boolean database. 3. Logical Data Analysis The logical analysis of data was originally developed for the analysis of datasets whose attributes take only binary (0-1) values [4, 5, 8].Since it turned out later that most of the real-life applications include attributes taking real values, a binarization" method was proposed in [3]. The purpose of binarization is the transformation of a database of any type into a Boolean database. Table 1.Original Database ID Age Age M.Status M.Status Single Married LAD is a methodology developed since the late eighties, aimed at discovering hidden structural information in Boolean databases. LAD was originally developed for analyzing binary data by using the theory of partially defined Boolean functions. An extension of LAD for the analysis of numerical data sets is achieved through the process of binarization consisting in the replacement of each numerical variable by binary indicator variables, each showing whether the value of the original variable is present or absent, or is above or below a certain level. LAD has been applied to numerous disciplines, e.g. economics and business, seismology, oil exploration, medicine etc. [15]. 3.1 Binarization The methodology of LAD is extended to the case of numerical data by a process called binarization, consisting in the transformation of numerical (real valued) data to binary (0, 1) ones. In this [7] transformation we map each observation u = (ua, ub, ) of the given numerical data set to a binary vector x(u) = (x1, x2, ) Є {0, 1}n by defining e.g. x1 = 1 iif ua α1, x2 = 1 iif ub α2, etc, and in such a way that if u and v represent, respectively, a positive and negative observation point, then x(u) x(v). The binary variables xi, i = 1,2,, n associated to the real attributes are called indicator variables, and the real parameters αi, i = 1, 2,, n used in the above process are called cut points. The basic idea of binarization is very simple. It consists in the introduction of several binary attributes associated to each of the numerical attributes; each of these binary attributes is supposed to take the value 1 (respectively, 0) if the numerical attribute to which it is associated takes values above (respectively, below) a certain threshold. Obviously the computational problem associated to binarization is to find a minimum number of such threshold values (cut points) which preserve the essential information contained in the dataset. In order to illustrate the binarization of business datasets, let us consider the examples presented in Table 1. A very simple binarization procedure is used for each variable age and marital status. Quantitative attributes such as age is divided into different ranges like age: , , etc. The marital status variable is divided into binary values by converting its domain values into attributes. Table 2 Boolean Database ID Age M.Status #cars 1 22 Single Married Binary Variables A binary variable has only two states: 0 or 1, where 0 means that the variable is absent, and 1 means that it is present. If all binary variables are thought of as having the same weight, we have the 2-by-2 contingency table of table 3, where a is the number of variables that equal 1 for both items i and j, b is the number of variables that equal 1 for item i but that are 0 for item j, c is the number of variables that equal 0 for item i but

314 292 equal 1 for item j, and d is the number of variables that equal 0 for both item i and j. The total number of variables is z, where z = a + b f(1,1, 1)/ f i10 i20 i 30, in 0 + c + d. f(i 1,i 2,..in) - f(0,0,..0). Eq. (3.3.1) Table 3. A contingency table for binary variables Item j b) If d < Φ then i 1,i 2,..in are frequent. Eq. (3.3.2) 1 0 Sum 1 A B a + b 0 C D c+ d Sum A + c b+ d Z For noninvariant similarities, the most wellknown coefficient is the Jaccard dissimilarity coefficient, where the number of negative matches d is considered unimportant and thus is ignored in the computation: Item i b c d( I, J ) Eq. (3.2.1) a b c The measurement value 1 suggests that the objects i and j are dissimilar and the measurement value 0 suggests that the objects are similar. This method is used in SI algorithm while the Apriori algorithm works using similarity measures. 3.3 SI Algorithm I use two algorithms for the implementation purpose. Our develop SI algorithm and the well known Apriori algorithm to check the accuracy and efficiency. Input Φ User specified threshold between 0 And 1 T Binary transactional Database Output Frequent itemsets Step p = { i 1,i 2,..in} set of data items in transactional database. Create K Map for all the permutation in row. Scan the transactional database and put the presence for every combination of data items in corresponding K Map for every permutation of row. For every permutation of p: a) Calculate dissimilarity using K Map Constructed for every permutation using the following Jacquard s dissimilarity equation. d (i1,i2,..in) = 1 i10 1 i20 0,0,..0) 1 i 30,. 1 in0 f(i 1,i 2,..in)f( 3.4 Apriori Algorithm Input Φ User specified threshold between 0 And 100 T Binary transactional Database Output Frequent lists Apriori algorithm work on similarity measure while the SI algorithm works on dissimilarity measure. 4. Experimental Results I performed different experiments to check the results and efficiency of the technique. The data required in database should be in binary format. I downloaded the dataset transa from the net [11]. The data was stored in a format: I coded the algorithms in ORACLE 10g using laptop computer having 20GB hard drive and 1.6MH processor. I create a table in the database to store the data for the purpose of experiment. To load the data to the database oracle provides a facility by making a control file and then by using SQL loader. We first convert the data into a format that the item now is separated by commas instead of spaces. Now the data is loaded to the table with the help of SQL loader and look like 0,1,0,0,1,0,0,1 After loading the data into table the algorithms are implemented on the database having fifty records, initially.

315 293 Figure 1.SI Algorithm Figure 2. Apriori Algorithm Figure 4.Apriori Algorithm results The largest frequent lists generated by the algorithm are I 1, I 2, I 3,, I 1, I 3, I 4 After giving the data to Apriori algorithm it also produce the same results. With the same largest frequent sets contain, I 1, I 2, I 3 I 1, I 3, I 4 After loading more data, the total records in the database became 500. Applying Apriori and SI algorithms on the updated database, the results produced are given. The largest frequent list. I 1, I 2, I 3 I 1, I 3, I 4 We found that again the algorithms produce the same results. After loading more data the total number of records become And again applying the algorithms on the database, the results produce are given below. I 1, I 2, I 3 I 1, I 3, I 4 So this is again that the algorithms produce the same results. After loading more data to the database the total records in the database are Again applying Apriori and SI algorithm on the database the results produced are given. I 1, I 2, I 3 I 1, I 3, I 4 Figure 3. SI Algorithm results We see that again the algorithms produce the same results. Now we have to load more data to the database the total number of records become And once again applying both the algorithm on the database the results produce are given below. I 1, I 2, I 3 I 1, I 3, I 4 We see that again the algorithm produces the same results. Up to this we analyze the performance, efficiency and accuracy of the algorithms by changing size of the database. And this is clear

316 294 that the results produce from this database are very consistent and reliable for the evaluation of learners, tutors and subjects etc. After this we change the input threshold to analyze the performance, efficiency and accuracy at different threshold values. The input threshold changes from.80% to.70% (dissimilarity) for SI algorithm and from 20% to 30% (similarity) for Apriori algorithm. The database contains 8000 records and after applying both the algorithms the results produces are given below. Figure 5 SI Algorithm Results 6 Concluding Remarks And Future Work. In this research, I study that how data mining techniques are used for the evaluation and Knowledge discovery in the field of education. The output produced was based on realistic reasons and values so it is reliable, efficient and precise for the experts. Here we produce the results by performing different experiments and prove that such techniques are very consistent. On the basis of the output we may predict and evaluate students and teachers. And also we take the results to make categories of the learners. By this we may enhance the learning process. The subjects may be divide into groups on the basis of similarity, dissimilarity measure. Further we may perform experiments for other algorithms from different point of view on different data storage. Figure 6. Apriori Algorithm Results The largest frequent list produce is I 1, I 3, I 4 Now this is clear that both algorithms produced the same results at different threshold. Up to this, we analyze that these techniques are very reliable for the analysis and discovery of hidden pattern and information in any type of database just like in this educational database. Time 5 Graphical analysis Data Figure 7. Graphical analysis of the results 7 References [1] Comer, D. E.; Gries, D., Mulder, M. C., Tucker, A., Turner, A. J., and Young, P. R. "Computing as a discipline". Communications of the ACM. (Jan. 1989) vol32 (1) [2] Wegner, P. "Research paradigms in computer science". Proceedings of the 2nd international Conference on Software Engineering. San Francisco, California, United States: Press, Los Alamitos, CA. (October 13 15, 1976). [3] Boros E., P.L. Hammer, T. Ibaraki, A. Kogan. Logical Analysis of Numerical Data. Mathematical Programming, (1997). 79: [4] Boros E., P.L. Hammer, T. Ibaraki, A. Kogan, E. Mayoraz, I. Muchnik. An Implementation of Logical Analysis of Data. IEEE Transactions on knowledge and Data Engineering, 12(2) ( 2000): [5] Crama Y., P.L. Hammer, T. Ibaraki. Cause-effect Relationships and Partially Defined Boolean Functions. Annals of Operations Research, 16(1988).: [6] David Wai-Lok Cheung, Vincent T. Ng, Ada Wai-Chee Fu, and Yongjian Fu.. Efficient Mining of Association Rules in Distributed Databases, IEEE Transactions on Knowledge and Data

317 295 Engineering, (December 1996)Vol. 8, No. 6, pp [7] E. Boros, P. L. Hammer, T. Ibaraki, A. Kogan, E. Mayoraz and I. Muchnik An implementation of logical analysis of data, RUTCOR Research Report RRR 22-96, Rutgers University, 1996.,pp [8] Hammer P.L. The Logic of Causeeffect Relationships, Lecture at the International Conference on Multi- Attribute Decision Making via Operations Research-based Expert Systems, (1986)Passau, Germany. [9] Hannu Toivonen. Sampling Large Databases for Association Rules, Proceedings of the 22nd International Conference on Very Large Databases, (1996) pp , Mumbai, India. [10] Heikki Mannila, Hannu Toivonen, and A. Inkeri Verkamo Efficient Algorithms for Discovering Association Rules, Proceedings of the AAAI Workshop on Knowledge Discovery in Databases (KDD-94), (July 1994).pp [11] 1/ notes/itemsets/item set prog1.htm [12] Irshad Ullah, Abdus Salam and Saifur-Rehman,Dissimilarity Based Mining for Finding Frequent itemsets. Proceedings of 4th international conference on statistical sciences (2008) Volume (15), University of Gujrat Pakistan, 15: 78 [13] M. Houtsmal and A.. Set-Oriented Mining for Association Rules in Relational Databases, Proceedings of the 11th IEEE International Conference on Data Engineering, Swami (1995)pp , Taipei,Taiwan. [14] Ming-Syan Chen, Jiawei Han and Philip S. Yu.. Data Mining: An Overview from a Database Perspective, IEEE transactions on Knowledge and Data Engineering, (1996) Vol. 8, No. 6, pp [15] Peter L. Hammer Tiberius Bonates.. Logical Analysis of Data: From Combinatorial Optimization to Medical Applications, RUTCOR Research Report( 2005) RRR [16] Rakesh Agrawal and Ramakrishnan. Fast Algorithms for Mining Association Rules in Large Databases, Proceedings of the Twentieth International Conference on Very Large Databases, Srikant.(1994)pp , Santiago, Chile. [17] R. Agrawal, T. Imielinski, A. Swami. Mining Associations between Sets of Items in Massive Databases, Proc. of the ACM-SIGMOD Int'l Conference on Management of Data,Washington D.C. (1993). [18] Ramakrishnan Srikant. Fast Algorithms for Mining Association Rules and Sequential Patterns,Ph. D. Dissertation,University of Wisconsin, Madison. (1996) [19] Usama M. Fayyad, Gregory Piatetsky- Shapiro, and Padhraic Smyth. From Data Mining to knowledge Discovery: An Overview, Advances in Knowledge Discovery and Data Mining, (1996). AAAI Press,, pp Author: Irshad Ullah MS-CS (Database systems) 2006, B.Ed 2004 M.SC Computer Science 2002 B.SC Computer Science 2000 Senior IT Teacher (SITT) Since Research Interest: Data Mining Member of the: ISOSS. Two publications and presentations an International Conferences.

318 ISSN (Online): WLAN Security Flaw: Cracking 64 bit WEP Key Anil Kumar Singh 1, Bharat Mishra 2 and Sandeep Singh 3 1 Jagran Institute of Management, Kanpur (India) 2 MGCGV, Chitrakoot Satna (M.P.) India 3 Jagran Institute of Management, Kanpur (India) Abstract We put on display an active attack on the WEP protocol that is able to recover a 64-bit WEP key using 5000 capture packets with a success probability of 90%. In order to succeed 100% in all cases, more than 5000 packets are needed. The IV of these packets can be randomly chosen. This is an improvement in the number of required frames by more than an order of level over the best known keyrecovery attacks for WEP. In this paper we demonstrate the security flaws of Wireless LAN by cracking 64 bit WEP key on Wi-Fi access points using Backtrack, a live Linux distribution. We attack the Wi-Fi AP, making it generate packets for our cracking effort, finally cracking the WEP key successfully. Key words - WLAN, Wi-Fi, MAC address, radio waves, WEP Key, Access Point and IV. Introduction Wired Equivalent Privacy (WEP) is used to keep wireless connections secure from sniffing attacks. You ve probably heard that it s not very secure. It is a protocol for encrypting wirelessly transmitted packets on IEEE802.11networks. In a WEP protected network, all data packets are encrypted using the stream cipher RC4 under a common key, the root key [1]. The root key is shared by all radio stations [2][1]. A successful recovery of this Key gives an attacker full access to the network. WEP is a part of the IEEE wireless standard ratified in 1999 [3]. It was designed to provide confidentiality on wireless communications by using RC4. In order to simplify the key setup, WEP uses preinstalled fixed keys. The first analysis of the WEP standard was done in 2001 by Borisov, Goldberg and Wagnerin [4]. They demonstrated major

319 ISSN (Online): security flaws revealing that WEP does not provide confidentiality, integrity and Access control. Material and methods The research was carried out to reveal WLAN Security Flaw: Cracking 64 bit WEP Key. The work was conducted at Department of Information Technology, Jagran Institute of Management, affiliated to GBT University, Uttar Pradesh. Details of materials used and the procedures employed are as follows: We can design a scenario after understanding the theory of WEP cracking. We have taken the cracking software Back Track 3.0 [5]. Hardwares picked for the purpose were: HCL Desktop, Compaq Laptops, AP (D-Link 3200 Series Access Point) and Wireless card (D-Link DWA 510); and softwares picked were: Operating System (Windows XP) and other application softwares. A client is used to communicate with Access Point while User Datagram Protocol (UDP) flooding is used to send data. Another client is used to keep track of the network traffic as a hacker and listens to the WLAN. AP is linked to LAN with wires. Figure 1 is the illustration of the WEP cracking job. bt ~ # airmon-ng MAC-Media Access Control address provided by the manufacturer at the time of manufacturing. It is a unique 48 bit, hexadecimal form, hardware address of WNIC. We here change the MAC address of WNIC. Before changing the MAC address it should be noticed that WNIC should be down. bt ~ # ifconfig wlan0 down bt ~ #macchanger mac 00:11:22:33:44:55 wlan0 UDP flooder simulating the traffic Wireless Network Access Point Cracker System equipped with Backtrack Wired Network Figure 2 showing the fake MAC address bt ~ #airmon-ng start wlan0 Find out the Access Point and their MAC addresses, Data, Channel, Speed in MB, Encryption, Cipher, Authentication and ESSID. We type the following command in console window. bt ~ #airodump-ng wlan0 Figure 1 WEP gears cracking lab setup After booting with Back Track CD we open the console window and we find out the Wireless Network Interface Card (WNIC).

320 ISSN (Online): aireplay-ng -b <MAC address of the Access Point> pokemon-01.cap [6]. Figure 3 showing the MAC of AP, their Channel No. Speed, Encryption, Cipher and ESSID bt ~ # airodump-ng -c <channel No.> -w [pokemon] --bssid <MAC address of the Access Point> After execution of the above command, the output data which will be stored in the file pokemon; you can put any name of the file as you wish. This file will be offered for the WEP Crack program when we are ready to crack the WEP key. Open another shell and put down the previous command running. Now we need to generate some fake packets to the access point to speed up the data output. Test the access point by issuing the following command: bt ~ #aireplay-ng a <MAC address of the AP> -h <Fake MAC address of the WLAN> WLAN0 Figure 5 showing the associated AP Figure - 6 showing the attack with IVs If the above command is successfully executed, then we will have to generate many packets on the target network so that we can crack the WEP Key. bt ~ #aireplay-ng -3 -b <MAC address of the AP> -h <Fake MAC address of the WLAN> WLAN0 Figure 7 showing the decrypted WEP key Experimental Results The number of remarkable packets capturing is normally distributed l as shown in Figure 3. First, the capturing number will increase and then it will decrease as time passes. Figure 4 showing the speed of capturing packets It will force the Access Point to send out a bunch of packets which we can then use to crack the WEP key. After about capturing the IVs We start cracking the WEP key by typing the following: When D-link DWL-3200 series AP as AP (Center Point) and D-Link DWA 550 Series Client as client, it will generate repeat IV Packets in 2-3 minutes. And Back Track can crack WEP key successfully every time within seconds.

321 ISSN (Online): The results of this experiment shows: First, the cracking will become easier if the number of IVs generated is more. Secondly, the experiment paves way for the security up gradation of D-Link AP (Wireless device). Conclusion WLANs are largely used in education, healthcare, financial industries, and various public places such as airline lounges, coffee shops, and libraries. Although the technology has been standardized for many years, providing wireless network security has become a critical area of concern. Due to the broadcast nature of the wireless communication, it becomes easy for an attacker to capture wireless communication or to disturb the normal operation of the network by injecting additional traffic. 5. Download backtrack 3.0 by /backtrack.html 6. Maz, 19 Aug. Tutorial: Cracking WEP Using Backtrack 3 7. Ye Peisong, and Yue Guangxue (2010), Security Research on WEP of WLAN, Proceedings of the Second International Symposium on Networking and Network Security (ISNNS 10) p.p Anil Kumar Singh, MCA Asst. Professor, Jagran Institute of Management, Kanpur. Currently pursuing the Doctoral programme in WLAN Security Vulnerability Threats and Alternative Solution. at MGCV Satna (M.P.) Dr. Bharat Mishra, Ph.D., Dept of Physical Science. MGCGV Satna (M.P.) Security is of ultimate importance to the global communication and information networks. The data, which are encrypted with WEP Key, are also insecure [7]. WLAN is also prone to an authorized intervention by hackers because of its weakness analysed above. Sandeep Singh, PGDCA, Dept. of Information technology, Jagran Institute of Management, Kanpur. References 1. Tews Erik et. al.,breaking104 bit WEP in less than 60 seconds, TU Darmstadt, FB Informatik Hochschulstrasse 10, Darmstadt, Germany. 2. Shaheen Jaleel et. al. (2007), Confidential and secure broadcast in wireless sensor networks, The18th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications. 3. IEEE: ANSI/IEEE standard802.11b: Wireless LAN Medium Access Control, (MAC) and Physical Layer (phy) Specifications (1999). 4. N. Borisov et. al. (2001) - Intercepting mobile communications the Insecurity of In MOBICOM

322 ISSN (Online): Digitizing the Forest Resource Map Using ArcGIS Mrs. K.R. Manjula 1, Dr. S. Jyothi 2 and Mr. S. Anand Kumar Varma 3 1 Associate Professor of MCA Dept., Siddhartha Institute of Engineering and Technology, Puttur, Andhra Pradesh, India. 2 Associate Professor & HOD, Computer Science Department, SPMVV (Women s University), Tirupathi, Andhra Pradesh, India 3 Associate Professor & HOD of Civil Engineering Dept., Siddhartha Institute of Engineering and Technology, Putter, Andhra Pradesh, India, Abstract The major challenges we face in the real world today is overpopulation, pollution, deforestation, natural disasters which have a critical geographic dimension and also have a geographical component. Geographical Decision Support System (Geo-DSS) is a demanding field, since enormous amount of spatial data have been collected in various applications, ranging form Remote Sensing to GIS, Computer Cartography, Environmental Assessment and Planning. Although some efforts were made to combine spatial mining with Spatial Decision Support System but mostly researchers are using a spatial database for popular data mining approach. GIS will give you the power to create maps, integrate information, visualize scenarios, solve complicated problems, present powerful ideas, and develop effective solutions like never before so that it supports strategic decision making. With GIS application we can open digital maps on computer, create new spatial information to add to a map, create printed maps customized to our needs and perform spatial analysis on it. There is a great deal of geographic data available in formats that can not be immediately integrated with other GIS data. In order to use these types of data in GIS it is necessary to align it with existing geographically referenced data. This process is also called georeferencing. Georeferencing is a necessary step in the digitizing process. Digitizing in GIS is the process of tracing, in a geographically correct way, information from images/maps. In this paper the processing of digitizing the forest map and the converting the georeferenced map into attribute data is depicted which can be further used to construct spatial database and on which spatial analysis can be performed. Keywords ArcGIS Desktop, Digitization, Georeferencing, Spatial Attributes, Attribute Table. 1. Introduction Data acquisition and data storage technology progress has led to a huge amount of data stored in repositories, which grow faster. Among increasing and relevant data acquired and processed, there is a strategic segment: satellite images, also known as Remote sensing images. The search for less expensive and more efficient ways to observe earth, motivated man in developing Remote Sensing satellites. The Remote Sensing image based databases are the fastest growing archives of spatial information. The variety of spatial and spectral resolutions for a Remote Sensing images range from IKONOS 1-meter panchromatic images to the next generation of polar-metric radar imagery satellites. There is a need for understanding relevant data and use it effectively and efficiently in earth analysis. Although valuable information is contained in image repositories, the volume and complexity of this data makes difficult (generally impossible) for human being to extract strategic information (knowledge) without appropriate tools (Piatetsky-Shapiro, Djeraba, Getoor, Grossman, Feldman & Zaki, 2006). The availability of huge Remote Sensing image repositories demands appropriate resources to explore this data. A vast Remote Sensing database is a collection of landscape snapshots, which supplies a single opportunity to understand how, when and where changes occurred in the world. The first operational remote sensing satellite (LANDSAT-1) was launched in the year1972. According to Canada Centre for Remote Sensing (2003), Remote Sensing is the science (and to some extent, art) of acquiring information about the Earth surface without actually being in contact with it. Remote Sensing is a field of applied sciences for information acquisition of the Earth surface through devices called Remote Sensors, which are boarded

323 ISSN (Online): on Remote Sensing aircrafts or satellites also called Earth observation satellites. In the image acquisition process, four concepts are fundamental: spatial, spectral, radiometric and temporal resolution. The spatial resolution defines the detail level of an image. The spectral resolution determines the sensor capability to define short intervals of wavelength. The radiometric resolution of an imaging system describes its ability to discriminate very slight differences in energy. The temporal resolution determines the necessary time for the sensor revisit a specific target and image the exact same area (Canada Centre for Remote Sensing, 2003). These huge volumes of the datasets need an efficient hardware and software infrastructure. However, still a limited capacity is available for extracting information from large remote sensing image databases. Currently, most image processing techniques are designed to operate on a single image. This situation has lead to a knowledge gap in the process of deriving information from images and digital maps (MacDonald, 2002). There are number of studies that shown the overlapping of various map to obtain the required information. The main aim of this paper is to depict the process of digitizing the forest resource map and convert it in to attribute data. The converted attribute data can be stored in to a database tables for future work. 2. Methodology In this paper an implementation procedure for the construction of a spatial attribute table from scratch by digitizing the forest area map and generating topologically corrected data set is derived. You are given a scanned map of forest area of AP. At first, taking the scanned map as a base map, you will digitize using ArcCatalog, and save the digitized data into attribute tables. The methodology for the above said task is depicted in the following sub sections. 2.1 Data Collection As the main aim of this paper is to depict the process of digitizing the forest resource map and convert it in to attribute data, the data collection phase derives data from various sources. The data are collected from divisional forest office, district land and land reform office etc. The maps may also be acquired from the FSI, NRSC and Google Earth Data Preprocessing Results from this project suggest that the use of digital image processing in conjunction with editing, georeferencing and spatial analysis in a Geographic Information System is an effective mean of quantifying deforestation, and that the use of high resolution LANDSAT data may in fact yield much better precision than AVHRR-based analyses. The use of digital preprocessing with visual post-processing greatly reduces analysis time over that of hand digitizing of a photographic product and greatly reduces the confusion of classes associated with purely digital processing techniques Data Editing This stage can further be subdivided into various sub stages, like Software: In this study the used software packages are ArcGIS Desktop. ArcGIS ArcGIS is an integrated collection of GIS software products for building a complete GIS. It consists of a number of frameworks for deploying GIS: ArcGIS Desktop an integrated suite of professional GIS applications ArcGIS Engine embeddable developer components for building custom GIS applications Server GIS ArcSDE, ArcIMS, and ArcGIS Server Mobile GIS ArcPad as well as ArcGIS Desktop DESKTOP GIS Desktop GIS is the primary seat from which GIS professionals compile, author, and use geographic information and knowledge. ArcGIS Desktop is an integrated suite of advanced GIS applications. It includes a series of Windows desktop applications ArcMap, ArcCatalog, ArcToolbox, and ArcGlobe with user interface components. Using these applications and interfaces in union, you can perform any GIS task, simple to advanced, including mapping, geographic analysis, data editing and compilation, data management, visualization, and geoprocessing. ArcGIS Desktop is scalable to meet the needs of many types of users. It is available at three functional levels: ArcView focuses on comprehensive data use, mapping, and analysis. ArcEditor adds advanced geographic editing and data creation ArcInfo is a complete, professional GIS desktop, containing comprehensive GIS functionality, including rich geoprocessing tools.

324 ISSN (Online): DIGITIZATION Digitizing is the process of converting features on a paper map into digital format. You can use a digitizer in conjunction with the editing tools in ArcMap to create new features or edit existing features on a digital map. Digitization is process of making an electronic version of a real world object or event, enabling the object to be stored, displayed and manipulated on a computer, and disseminated over networks and/or the www. But before digitization, the map must be georeferenced. GEOREFERENCING There is a great deal of geographic data available in formats that can not be immediately integrated with other GIS data. In order to use these types of data in GIS it is necessary to align it with existing geographically referenced data, called georeferencing. The process of georeferencing relies on the coordination of points on the scanned image (data to be georeferenced) with points on a geographically referenced data (data to which the image will be georeferenced). By linking points on the image with those same locations in the geographically referenced data you will create a polynomial transformation that converts the location of the entire image to the correct geographic location. We can call the linked points on each data layer as control points. The selection of control points is important. DATA A.jpg file format can be used as data for georeferencing, which contains Andhra Pradesh forest map. After performing georeferencing, the file is saved as.mxd format. GEOREFERENCING/REGISTERING AN IMAGE PROCEDURE You should be using the clipped data for linking, and use the labels too. To add a link, click the mouse on the image first, then on a known location on the GIS database. Once you ve added a few links you can open the link table to see how each has worked. In the link table you can view the Total RMS Error as well as the residual for each link that contributes to the overall total control points. The errors are associated with the amount of disagreement between the two control points for each link once the transformation is set. You have to be sure that the two points that are linked are the someplace in world, as you could have a very low error and residuals but not have an accurately registered image. Once you are satisfied with the registration, and you can delete links or start all over by deleting them all, you can Update Georeferencing, this will save your transformation. This process will add a.aux file that contains information about the transformation that is necessary to view this file in the future with other data. Save your map as apf1.mxd. DIGITIZING FEATURES Digitizing is the process of making features. We can see on the AP forest image editable and making them features to which additional spatial and non-spatial attributes can be assigned. This means we are going to follow a process of making digital versions of objects that will have an attribute table with them. Once they are digitized and have an associated attribute table these objects will also be known as polygon features. By digitizing these features you make them available for mapping once you have added the tabular data to the attributed table. The digitizing process is started by creating new layers in ArcCatalog, and then adding features in ArcMap. The Primary tool used is the georeferencing toolbar. Add the APforest.jpg file. Right click the file (not image) and Zoom to Layer. From the georeferencing toolbar, click the Layer dropdown arrow and select the APforest.jpg image name. Click Georeference and select Fit to display from the dropdown. Now, it is ready to add control points. Click on the control point s tool. CREATION OF NEW LAYER BY CLIPPING Open the layer (AP forest map). Open Arc Toolbox (red toolbox on menu) and open Analysis Tools, then Extract. Choose the clip tool. The input feature is the forest map. The clip feature is the AP forest map. The output class should go somewhere you can find it and be named something you will recognize.

325 ISSN (Online): Left the cluster tolerances as it is and Click OK. Add the new layer and remove the original map layer. CREATING AN EMPTY SHAPE FILE In ArcCatalog, browse to the location of your current.mxd file. This is the folder in which you will create your new shape file, so select that folder and right-click on it. Go to New and select shapefile and give an appropriate name such as Area. Click on Edit to see the coordinate system of the file. In the Spatial Reference Properties window click Import to use the projection of the AP forest layer. Click OK and OK again to create the shapefile. ADD A NEW FILED IN THE ATTRIBUTE TABLE Return to ArcMap, and add your new shapefile to the Data frame (TOC). If you open the Attribute Tables of this shapefile you will find it empty. Now you will start working with the new shapefile. Before you start editing, first open its Attribute table. Click on the Options Button and click on Add Field. First create a Short Integer (choose in the Type drop-down menu) field called Area where the area number identifier will be entered. Note that you cannot create new fields while you are editing a layer. DIGITIZING AREAS AND ENTERING TABULAR DATA You will need the Editor Toolbar. Select View then select Toolbars and then select Editor. On the toolbar, click on the Editor Menu and Start Editing. You will be prompted to choose the folder that your shapefiles are in, and then click OK. Before you get started creating polygons, turn off all but the registered forest image layer. In the Editor Toolbar, start Editing. Choose Area (whatever you named it), layer as the layer you want to edit. In the Editor Toolbar, Create New Feature. You are going to create a polygon for maps. From the pull down menu next to the editor pull down choose the sketch tool. By clicking at the four corners (Single clicks) of the area you are digitizing you will create a continuous outline of it. Double clicking when you get back to where you started you will finish Sketch. This should make your polygon become an actual, filled polygon. You can digitize 3-5 area boundaries and enter corresponding area ID numbers into the AREA field in the attribute table. Now that you have digitized a few polygons in ArcGIS you have acquired the skills to create editable features from scanned maps that can be linked with tabular data. If you are interested you can do a number of things to extend what you have learned. DATABASE CONSTRUCTION PROCEDURE Part B, of this project is intended to instruct you on how to integrate the Microsoft Access into Arc View using the Andhra Pradesh environmental sampling data as an example. Then it list the procedure for how to set up the link between Access and Arc View, by which you can use Access to store and manage the environmental sampling data, and use Arc View to display and analyze the subset data of interest. 3. Review of Literature In Analyzing deforestation rates, spatial forest cover changes and identifying critical areas of forest cover changes in North-East India during ,by Nikhil Lele P. K. Joshi - The objective of this paper is to depict forest cover change (namely deforestation) processes occurring in North-East India, with the aid of geographic information system (GIS). The results are aimed at understanding how GIS methods can provide rapid and precise outputs that may help to improve conservation policies and land use-planning strategies. The temporal forest cover datasets were acquired from the available sources. NRSA, India made use of satellite datasets and mapped forest cover of India in 1972; followed by a second time mapping in This mapping was carried out at a coarse scale of 1:1M, using Landsat datasets. The maps were thus converted into digital database and imported into geospatial environment using ERDAS Imagine 8.7 and ArcGIS 8.3. In Monitoring and Mapping India s Forest and Tree Cover through Remote Sensing, by J. K. Rawata, Alok Saxenab and Saibal Dasguptac - In its latest assessment of 2001, taking advantage of advancements in remote sensing and improvement in digital interpretation qualities, FSI has provided a much more comprehensive status of forest cover in the country. In Detecting Tropical Deforestation Using Satellite Radar Data, by Belinda Arunarwati and Yousif Ali Hussin - The main objectives of this research were to investigate the potential of satellite radar data for detecting, differentiating and classifying deforestation in the forest concession area which has been selectively logged by PT Sylva Gama, Jambi, Central Sumatra, Indonesia. The following satellite images were used for this research project: Landsat-5 TM data of September 15, 1993, Spot XS data of March 21, 1993, ERS-1 images of October 17, 1993, June 6, 1994, and July 7, 1994, and JERS-1 of August 16, Conclusions The power of a GIS as an aid in spatial data analysis lies in its geo-relational Database structure, i.e., in the Combination of value information and location

326 ISSN (Online): information. The link between these two allows for the fast computation of various Characteristics of the spatial arrangement of the data, such as the contiguity structure between observations, which are essential inputs into spatial data analysis. The GIS also provides a flexible means to "create new data," i.e., to transform data between different spatial Scales of observation, and to carry out aggregation, partitioning, interpolation, overlay and buffering operations. Andhra Pradesh has experienced a combination of land cover changes such as deforestation and afforestation since past 30 years. For any spatial analysis, the Pre-dominant step is Digitization. So in this process, digitizing the map features of Andhra Pradesh by using GIS software called ArcGIS is depicted. By digitizing the map, it is clear and make easy for performing spatial analysis. The result of Digitization is more accurate and provides more information than Digitizing Tablet. In further enhancement, this work can also be carried out by any other digitizing software like GRASS, Enguage Digitizer etc. This work also proceeds by classification to get clearer classified digitization map. Regional Land Use: The Clue-S Model, Environmental Management Vol. 30, No. 3, Pp [6] Quick Guide: Digitizing (Editing) GIS Map Layers. [7] Stefan Erasmi, André Twele, Muhammad Ardiansyah, Adam Malik And Martin Kappas, Mapping Deforestation And Land Cover Conversion At The Rainforest Margin In Central Sulawesi, Indonesia, Earsel Eproceedings 3, 3/2004. [8] Velázquez, A., E. Durán, I. Ramírez, J.-F. Mas, G. Bocco, G. Ramírez, and J.-L. Palacio Land use-cover change processes in highly biodiverse areas: the case of Oaxaca, Mexico. Global Environmental Change 13(3): [9] Ye Maggie Ruan and David Maidment, Integrating Microsoft Access to ArcView, Center for Research in Water Resources The University of Texas at Austin. [10] Yuan Zhanliang *, Li Aiguo And Yie Ting, Research On Date Collecting And Rapid Map Updating Of Forest Resources Based On 3s Technology School Of Surveying And Mapping, Henan Polytechnic University, Jiaozuo , China The main aim of this paper is to depict the process of digitizing and converting the forest resource map in to attribute data. The converted attribute data is stored in to a database for future analysis which can be achieved in future work. Refrences [1] Georeferencing and Digitizing Image/Map, Spatial Structures in the Social Sciences Image/Map Georeferencing and Digitizing. [2] Kamal A. Niknami and A. Chaychi Amirkhiz, A GIS Technical Approach To The Spatial Pattern Recognition Of Archaeological Site Distributions On The Eastern Shores Of Lake Urmia, Northwestern Iran. [3] Lilian Pintea, Marvin E. Bauer, Paul V. Bolstad, Anne Pusey, Matching Multiscale Remote Sensing Data To Interdisciplinary Conservation Needs: The Case Of Chimpanzees In Western Tanzania, Pecora 15/Land Satellite Information Iv/Isprs Commission I/Fieos 2002 Conference Proceedings. 1. Mrs. K.R. Manjula, currently working as Associate Professor in MCA Department with 12 years of experience. The authour is also member of ISTE and CSI and also acting as Faculty Advisor for ISTE Student Chapter. She has presented her papers in more than 14 conferences and aattendended various workshops. She has guided more than 90 projects. 2. Dr. S. Jyothi, currently working as Associate Professor and heading the Department of Computer Science in Sri Padmavati Mahila Visva Vidhyalayam, Tirupati. She has more than 30 years of experience guided more than 100 PG, M.Phil and Ph.D scholars. She is meber of ISTE and CSI and IEEE. 3. Mr. S. Anand Kumar Varma, currenntly working as Associate Professor in Civil Engineering Department and also heading the Department and Exam Section in the SIETK. He has more than 15 years of experience both in industrey and academic and has done number of projects. He is life memer of ISTE. [4] Marcelino Pereira Dos Santos Silva, Maria Isabel Sobral Escada And Ricardo Cartaxo Modestode Souza, Remotesensing image mining: detecting agents of land-use change in tropical forest areas, International Journal of Remote Sensing Vol. 29, No. 16, 20 August 2008, [5] Peter H. Verburg* Sharifah S.A. Mastura Ramil Limpiada Victoria Espaldon, Modeling The Spatial Dynamics Of

327 ISSN (Online): Annexure Fig. 5 Arc Catalog Window Fig. 1 Arc Editor Window Fig. 6 Viewing an.mxd file format Fig. 2 Importing the map for Georeferencing Fig. 7 Creation of Shapefile Fig. 3 Physical Map of Andhra Pradesh Fig. 8 Spatial Reference Properties Window Fig. 4 Viewing the Link Table Fig. 9 Viewing the Shapefile in an Arc Catalog

328 ISSN (Online): Fig. 10 Attributes of Shapefile Fig. 11 Values of Attributes of Shapefile

329 Digital Watermarking in Discrete Wavelet Transformation - Survey 307 Rama Seshagiri Rao C. 1, Sampath Kumar M. 2 and Prakasam T. 3 1 Professor, CSE Department, Geethanjali College of Engineering and Technology, Cheeryal(V), Keesra(M), R.R.Dist., A.P , INDIA 2 Assoc. Professor, CSE Department, Geethanjali College of Engineering and Technology, Cheeryal(V), Keesra(M), R.R.Dist., A.P , INDIA 3 Professor, IT Department, Geethanjali College of Engineering and Technology, Cheeryal(V), Keesra(M), R.R.Dist., A.P , INDIA Abstract In the early days, encryption and control access techniques were used to protect the ownership of media. Recently, the watermark techniques are utilized to keep the copyright of media. Digital watermarking, a concept of embedding a special pattern, uses an algorithm of inserting a watermark to protect the copyright of media. It has recently become important in various application areas. The various of watermark techniques have been proposed by many authors in the last several years. However, there are not enough analysis and comparison on the previous researches. In this paper, for the understanding of the previous works and for the future related research, we try to classify and analyze the conventional watermark techniques from the various points of view. Currently watermark techniques based on the transfer domain, are more popular than those of the spatial domain. DCTbased methods have been most widely used among the transform based methods. However, recently wavelet based watermark techniques are becoming main research topic. With wide use of internet, effective audio and video watermarking researches are also required. Keywords: Digital Watermarking, encryption techniques, wavelet transformation, Image Processing. 1. Introduction Data is getting distributed rapidly in the world via the internet. It is very much required to authenticate the data to whom it belongs to. This has increased the requirement of the reliable and secure copy right protection technique invention. Digital watermarking and fingerprinting are the techniques in the field of Digital Rights Management (DRM). The companies working on tools for digital watermarking and fingerprinting are expecting the market growth to reach more than US$500 million worldwide by 2012, according to a new Multimedia Intelligence report[1]. Companies, such as Cinea, Philips and Verimatrix, are positioning transactional digital watermarking for integration into settop boxes, to increase content traceability and security. Fox Company has already announced that for early-release of high definition content, watermarking shall be made a compulsory component. Digital watermarking is a technique to embed copy right protection signature within multimedia contents. Watermark is nothing but an embedded signature, which can be an image or any type of media. A robust watermarking method provides a mark that cannot be removed from the watermarked content. For a watermarking technique to be robust, the watermark should be embedded in the perceptually significant portion of the data. Some typical distortions or attacks that digital watermarking schemes are expected to survive, include resampling, rescaling, compression, linear and nonlinear filtering, additive noise, A/D and D/A conversion, and transcoding. 2. Classification of Digital Watermarking Techniques There are many watermark techniques in terms of their application areas and purposes. And they have different insertion and extraction methods. The following table demonstrates the watermarking classification.

330 308 Table 1. Classification of watermarking according to several view points Classification Inserted Media Category Perceptivity of watermark Robustness of watermark Inserting watermark type Processing method in Spatial Domain Transform Domain Necessary Data for extraction 3. Background on wavelets Contents Text, image, audio, video visible, invisible robust, semi-fragile, fragile Noise, image format LSB, patch work, random function Look-up table, spread spectrum Private, semi- private, public watermarking Wavelets are functions that satisfy certain mathematical requirements and are used in representing data or other functions. Approximation using superposition of functions has existed since early 1800 s, when Joseph Fourier discovered that he could superpose sine and cosines to represent other functions. Wavelet algorithms process data at different scales and resolutions i.e. the resultant activity depends on the perception of the requirement. There are various wavelets: Haar, Coiflet, Daubechie, etc. Whereas the basis function of the Fourier transform is a sinusoid, the dyadic wavelet basis is a set of function which are defined by a recursive difference equation M-1 Ф(X) = CK Ф(2x-k) K=0 where M is the number of nonzero coefficients. The value of coefficients is determined by constraints of orthogonality and normalization. Wavelet transform uses wavelets as basis and is a tool that cuts up data or functions or operation into different frequency components, and then studies each component with a resolution matched to its scale. 3.1 Wavelet based Watermarking Watermarking in the Discrete Wavelet Transform (DWT) domain consists of encoding and decoding parts. In the encoding part, we first decompose an image into wavelet frequency domain to obtain decomposed image. Image permuted watermark we add to obtained image decomposition. The watermark permutation is reversible and it is the key for correct watermark extraction. For each coefficient within the wavelet domain, the key has a corresponding value of one or zero (if watermark is a binary image) to indicate if the coefficient is to modify or not. Note that watermarks are not inserted into the LH1, HL1 and HH1 bands (where L denotes the low pass band and H denotes the high pass band), because the energies in these bands are relatively small. In decoding part, we then take the twodimensional (2D) inverse DWT (IDWT), obtaining the watermarked image I. 4.0 Watermarking Algorithms in DWT Domain: Zhu and Tewfik [2] proposed two techniques, which uses a mask in the spatial or frequency domain. Generally, the effects of space or frequency masking are often used to form sequences of pseudo noise in order to maximize the energy of a watermark while maintaining the watermark itself invisible. The authors used the masking values obtained by the model with visual threshold from their work on image binary rate low coding. Inoue [3] invented a method that applies a DWT to the image, inserts the watermark in low frequency Sub bands and carries out an inverse DWT to obtain the watermarked image. This method allows the embedding of the watermark bits in the same block from which they were extracted, which help enabling good detection and localization of corrupted regions. The obtained watermark is then compared with the bits extracted for each block. If the number of different bits exceeds a predefined threshold, the corresponding block is considered altered. Otherwise the block is authentic. Paquet and Ward proposed a method based on the DWT [4]. This method was also simulated by Kundur and Hatzinakos with some modifications[5]. The proposed algorithm can detect and localize any tampering with acceptable precision. It is robust against JPEG2000 compression that is based on DWT. Moreover, its security is based on the key security that must be transmitted separately through a secure channel. In [6], S.C.Chu, H.C. Huang, Y.Shi, S.Y.Wu,C.S.Shieh have used genetic algorithm to select appropriate zero trees in the wavelet transform to pursue both the watermarked image quality and the robustness of the extracted watermark under planned attacks. In [7], Angela D Angelo, Mauro Barni and Neri Merhav have tested the perceptual intrusiveness and desynchronization efficacy through generating more powerful classes of geometric attacks on images. In [8], X.B.Wen, H.Zhang, X.Q.Xu,J.J.Quan, the algorithm takes the host image and divides the image into small blocks and then by verifying whether the block can be used for embedding the watermark or not, a decision is taken for embedding watermark. This is mainly to avoid the perceptual degradation of the image. And trained probabilistic Neural networks is applied to recover the watermark. In[9], A.E.Hassanien, A.Abraham and C.Grosan, have used pulse coupled Neural Networks to enhance the contrast of the human iris image and adjust the intensity with the median filter. They then used PCNN segmentation algorithm to identifying

331 309 boundaries of the iris image and then used texture segmentation algorithm to isolate the iris from human eye. They then extracted texture feature from quad tree wavelet and then Fuzzy C-Means algorithm is applied to the quad tree for further processing of boundaries. And then iris codes are extracted that characterizes human iris by using wavelet theory. These codes are embedded in to host images to identify the owner. In the authentication process, hamming distance metric that measure the recorded iris code and the extracted code from watermarked image is used to test whether the image is modified or not. Lin and Delp proposed a method [10], to generate a smooth watermark that resists to damages caused by JPEG compression. Gaussian distributions of pseudo-random numbers, with a zero average and unit variance, are used to generate the watermark. In consequence, each block contains a different watermark, but the distribution of the watermark in all blocks is similar. Detecting the presence of the watermark is based on the difference between adjacent pixels in the spatial domain. Block s authenticity is decided by comparing the difference with a predefined threshold. The detection and localization capabilities of this algorithm are very acceptable, but its performance could be affected by the block size. 5.0 Conclusion Many watermarking techniques have been proposed by various authors in the last several years. In this paper, we tried to classify and analyze many previous watermarking methods for understanding them and a help for new researchers in related areas. We classified the previous works from the various points of view: the inserted media category, the perceptivity, the robustness, the inserting watermark type, the processing method and the necessary data for the watermark extraction. Most of researches handled the watermark techniques on image media. Invisible watermarking, robust watermarking and noise style embedding have been main issues in the previous researches. In terms of processing domain, transform domain has been used rather than the spatial domain. Especially DCT-based approach has been widely used among the transform domain approaches, however, currently wavelet-based approach which has the multiresolution characteristic, is getting its popularity day by day. With the broad spreading of internet, audio and video based services such as MP3 and VOD are also being widely used. Therefore, proper audio and video watermarking techniques are also required to study intensively. Acknowledgments We would like to convey our regards to Dr. M.V.N.K.Prasad for guiding us in pursuing our PhD programs. We also convey our regards to the Chairman and Principal, Geethanjali College of Engineering and Technology, for supporting by all means. References [1] Securitypark.net [2]. Zhu B, Swanson MD, Tewfik AH (1996). Transparent robust authentication and distortion measurement technique for images. IEEE digital signal processing workshop (DSP 96) [3]. Inoue H, Miyazaki A, Katsura T (2000) A digital watermark for images using the wavelet transform. Integr Comput Aided Eng 7(2): [4]. Paquet AH, Ward RK (2002) Wavelet-based digital watermarking for image authentication. In:Proceedings of the IEEE Canadian conference on electrical and computer engineering, vol I. Winnipeg, Manitoba, Canada, pp [5]. Kundur D, Hatzinakos D (1999) Digital watermarking for telltale tamper proofing and authentication. Proc IEEE 87(7): [6]. S.C.Chu, H.C. Huang, Y.Shi, S.Y.Wu,C.S.Shieh: Genetic watermarking for Zerotree based applications: Circuits Syst Signal Process(2008) 27: [7]. Angela D Angelo, Mauro Barni and Neri Merhav: Stochastic Image Warping for improved watermark desyncronization: EUROSIP Journal on Information Security:Vol.2008 Art.ID [8]. X.B.Wen, H.Zhang, X.Q.Xu,J.J.Quan: A new watermarking approach based on probabilistic neural network in wavelet Springer -Verlog 2008 [9]. A.E.Hassanien, A.Abraham and C.Grosan: Spiking neural networks and wavelets for hiding iris data in digital images, Published online: 7 June 2008, Springer-Verlag 2008 [10]. Lin ET, Christine I, Podilchuk B, Delp EJ (2000) Detection of image alterations using semi-fragile watermarks. In: Proceedings of the SPIE international conference on security and watermarking of multimedia contents II, vol 3971, San Jose, CA, USA [11]. Adil Haouzia & Rita Noumeir : Methods for image authentication: a survey; Published online: 1 August 2007 Springer Science + Business Media, LLC 2007 C.Rama Seshagiri Rao B.E.(1991), M.I.T.(1998), M.Tech(2008), pursuing PhD at JNTU Hyderabad. He is presently working as a Professor in Geethanjali College of Engineering and Technology. He is a Life member of CSI and ISTE chapters. He has published many papers in International Journals. He is currently pursuing his PhD in Digital Watermarking under the guidance of Dr.M.V.N.K.Prasad. M.Sampath Kumar B.E.(1998), M.Tech(2000). He is presently working as a Associate Professor in Geethanjali College of Engineering and Technology. He is a Life member of ISTE chapter. He has published many papers in International Journals. His research interest includes Image Processing, Remote Sensing and GIS applications. T. Prakasam B.Tech.(1991), M.Tech(1999). He is presently working as a Professor in Geethanjali College of Engineering and Technology. He is a Life member of ISTE chapter. He has published many papers in International Journals. His research interest includes Image Processing, Software Engineering and Data Base Management Systems.

332 310 A Methodology for Aiding Investment Decision between Assets in Stock Markets Using Artificial Neural Network PN Kumar 1, Rahul Seshadri.G 2, Hariharan.A 3, VP Mohandas 4, P Balasubramanian 5 1,2,3 Dept. of CSE, 4 Dept. of ECE 5 Amrita School of Business Amrita Vishwa Vidyapeetham, Ettimadai, Coimbatore, Tamil Nadu, , India Abstract This paper outlines a methodology for aiding the decision making process for investment between two financial market assets (eg a risky asset versus a risk-free asset), using neural network architecture. A Feed Forward Neural Network (FFNN) and a Radial Basis Function (RBF) Network have been evaluated. The model is employed for arriving at a decision as to where to invest in the next time step, given data from the current time step. The time step could be chosen on daily/weekly/monthly basis, based on the investment requirement. In this study, the FFNN has yielded good results over RBF. Consequently the FFNN developed enable us make a decision on investment in the next time step between a risky asset (eg the BSE Sensex itself or a single share) versus a riskfree asset (eg Securities like Govt Bonds, Public Provident Funds etc).the FFNN is trained with a set of data which helps in understanding the market behaviour. The input parameters or the information set consisting of six items is arrived at by considering important empirical features acting on real markets. These are designed to allow both passive and active, fundamental and technical trading strategies, and combinations of these. Using just six items simplifies the decision making process by extracting potentially useful information from the large quantity of historic data. The prediction made by the FFNN model has been validated from the actual market data. This model can be further extended to choose between any two categories of assets whose historical data is available. Keywords: financial forecasting, risky assets, risk free assets, Feed Forward Neural Networks, Radial Basis Function Networks. 1. Introduction Artificial Neural Networks (ANN) have earned themselves a unique position as non-linear approximators. In general, of all the AI techniques available, ANN deal best with uncertainty[1]. Like other forms of soft computing, ANN perform well in noisy data environments and has proved to exhibit a high tolerance to imprecision. These characteristics of ANN make them particularly suited to the arena of financial trading. The stock market represents a data source with an abundance of uncertainty and noise. While ANN have been extensively studied [2,3] to perform a predictive analysis of the security prices, it may not be possible to model one general network that will fit every market and every security, hence models are built specific to markets and asset classes. Risky assets are those which do not have a guaranteed rate of return. An example of risky asset is stocks. Risk-free assets are those which give a return at a constant rate for eg, securities like Govt bonds. Since there is always a particular amount of risk associated with a risky asset, an investor cannot be assured of making a profit. This creates a need for a choice between risky and risk-free assets in order to maximize the profits, while minimizing the risk at the same time. The historical market prices would provide us with a better idea about the market s behavior. This data is incorporated into a Feed-Forward Neural Network in order to predict the behavior of the market in the future. These have been modeled on the Bombay Stock Exchange (BSE) and the results are presented here. 2. Investment Decision : Risky Vs Risk Free Asset 2.1 Artificial Neural Network. An Artificial Neural Network (ANN) is a mathematical model or computational model that attempts to simulate the structure and/or functional aspects of biological neural networks. There are different types of Neural Networks. There are no standard rules available for determining the appropriate number of hidden layers and hidden neurons per layer. Smaller number of hidden nodes and hidden layers would render better generalization. A pyramid topology, which can be used to infer approximate numbers of hidden layers and hidden neurons has been suggested by Shih [4]. Azoff [5] suggests that a network with one hidden layer and 2N + 1 hidden neurons is sufficient for N inputs, and states that the optimum number of hidden neurons and hidden layers is highly

333 311 problem dependant. Gately[6] suggests setting the number of hidden nodes to be equal to the total of the number of inputs and outputs. Some researchers suggest training a great number of ANN with different configurations, and then select that configuration that performed best- Kim at al.[7] Finally, another reasonably popular method is used by some researchers such as Kim& Lee[8] and Versace et al[9], whereby genetic algorithms are used to select between possible networks given choices such as network type, architecture, activation functions, input selection and preprocessing. Another method Tan[1], starts with a small number of hidden neurons and increase the number of hidden neurons gradually. A detailed comparative study can be seen at Vanstone[10]. For the purpose of this study, modeling has been attempted using a Feed-Forward Neural Network (FFNN) and a Radial Basis Function Network (RBF) Modeling BSE Sensex BSE Sensex, the most popular Indian stock index has been chosen for the study. The time step considered here is one month. BSE Sensex Index data pertaining to trading months starting from the year 2003 to 2008, is used for training the network. The duration is long enough and covers adequate market fluctuation. The one year monthly closing values of the Sensex for the year 2009 are used as the validation data set Model Architecture of FFNN The functional form used is a FFNN (see Fig 1) with a single hidden unit with restricted inputs giving an output LeBaron[11,12,13]. The output is a simple function α(z t, w j ). The equations given below define the network, h k = g 1 (w 0,k z t, k + w 1,k ) (1) 6 α(z t ) = g 2 (w 2 + Σ w 3,k h k ) (2) k=1 g 1 (x) = tanh(x) (3) Fig. 1 FFNN: Decision between Risky Vs Risk Free Asset. 2.4 Input Values to the FFNN:The Information Set The information set consists of six items. These are designed to allow both passive and active, fundamental and technical trading strategies, and combinations of these [11,12]. Using just six items simplifies the decision making process by extracting potentially useful information from the large quantity of historic data The first three inputs are the returns on equity in the previous three time-steps, useful for technical trading. The fourth is a measure of how the current price differs from the rational-expectations price. The last two inputs measure the ratio between the current price and exponentially weighted moving averages of the price. Information set being: z t,1 = r t = log( (p t +d t )/p t-1 ) z t,2 = r t-1 z t,3 = r t-2 z t,4 = log(r p t / d t ) z t,5 = log(p t / m 1,t ) z t,6 = log(p t / m 2,t ) Where p t is the share price, d t is the dividend paid, r is a constant and m i,t is the moving average given by m i,t = ρ i m i,t-1 + (1- ρ i ) p t (5) with ρ 1 = 0.8 and ρ 2 = Results: Training & Testing FFNN g 2 (x) = ½(1+tanh(x/2)) (4) where z t is time t information and w j are parameters. k takes values from 1 to 6 so that the weight array {w} consists of 19 parameters. The output from the intermediate neuron k is denoted h k. The output from the network, α is a 0 or 1 which would suggest where to invest in the next time-step, a 0 indicating that risk-free asset would give higher returns for the particular time-step and a 1 indicating that investing in an Index Fund tracking the BSE would render higher returns. The MATLAB Neural Network Toolbox has been chosen for creating, training and testing the network. The FFNN was trained with inputs from historical prices of BSE index, calculated taking monthly closing prices of BSE stock index from the year 2003 to The network was tested with data pertaining to the year 2009 and the results have been found to validate the market scenario. It has been found that the FFNN with one hidden layer with six neurons has produced quite accurate results. The network prediction matched with the test data of 2009 market. The neural network thus establishes the functional dependency between the input parameters and the market behavior.

334 Radial Basis Function (RBF) network Radial Basis Function (RBF) network(fig 2) is an artificial neural network that uses radial basis functions as activation functions. It is a linear combination of radial basis functions. They are used in function approximation, time series prediction, and control. RBF networks typically have three layers: an input layer, a hidden layer with a non-linear RBF activation function and a linear output layer. The output, network is thus where N is the number of neurons in the hidden layer, is the center vector for neuron i, and, of the a i are the weights of the linear output neuron. In the basic form all inputs are connected to each hidden neuron. The norm is typically taken to be the Euclidean distance and the basis function is taken to be Gaussian. weights ai, Ci and β are determined in a manner that optimizes the fit between and the data. 2.7 Results: Training & Testing: RBF Network The MATLAB Neural Network Toolbox has been chosen for creating, training and testing the RBF network as well. The network was trained with inputs from historical prices of BSE index, as was done above, taking monthly closing prices during the period 2003 to The network was tested with data pertaining to the year However, the RBF network did not yield good results. This perhaps is attributable to the fact that the data presented to the network was close to each other, thereby resulting in improper clustering and inaccurate results. 3. Future Extension The model presented above can be extended to choose between any two risky assets. This strategy can be effectively employed to compare between two mutual funds, two index funds or to arrive at an investment decision between two stock exchanges even. Further studies can explore the strategy for comparison of more than two risky assets. Whereas in FFNN( in the case of Risky Vs Risk-free asset), the model only suggests a choice of where to invest, a possible extension might be to find out the proportion of wealth to be invested in Risky and Risk-free assets respectively. 4. Conclusion Fig 2 : Architecture of a radial basis function network. An input vector x is used as input to all radial basis functions, each with different parameters. The output of the network is a linear combination of the outputs from radial basis functions. The Gaussian basis functions are local in the sense that i.e. changing parameters of one neuron has only a small effect for input values that are far away from the center of that neuron. RBF networks are universal approximators on a compact subset of. This means that a RBF network with enough hidden neurons can approximate any continuous function with arbitrary precision. The Two neural network models, FFNN and RBF have been designed, tested and validated from the BSE data to enable an investor make a decision at different time steps. FFNN with one hidden layer with six neurons has produced quite accurate results. However, the analysis carried out on a RBF network did not yield good results. The model of FFNN pertains to arriving at a decision between investment in a risky and a risk-free asset. Hence the models suggested by this paper can be used as a tool for an informed investment decision in the share markets. It is hoped that this can bring about a better investment strategy and help in achieving greater profits to investors. References [1].Tan, C.N.W., Artificial Neural Networks: Applications in Financial Distress Prediction and Foreign Exchange Trading (Gold Coast,QLD: Wilberto Press, 2001). [2].White, H., Economic Prediction using Neural Networks: The case of IBM Daily Stock Returns. Second Annual IEEE Conference on Neural Networks 1988, [3].Oppenheimer, H.R. and G.G. Schlarbaum, Investing with Ben Graham: An Ex Ante Test of the Efficient Markets

335 313 Hypothesis, Journal of Financial and Quantitative Analysis, XVI(3), 1981, [4].Shih, Y., Neuralyst Users Guide: Cheshire Engineering Corporation, [5].Azoff, M.E., Neural Network Time Series Forecasting of Financial Markets (Chichester: Wiley, 1994). [6].Gately, E., Neural Networks for Financial Forecasting (New York: Wiley, 1996). [7].Kim, J.-H., et al., Stock Price Prediction using Back propagation Neural Network in KOSPI. International [8].Kim, K.-J. and W.B. Lee, Stock market prediction using artificial neural networks with optimal feature transformation, Neural Computing and Applications, 13(3), 2004, [9].Versace, M., et al., Predicting the exchange traded fund DIA with a combination of genetic algorithms and neural networks, Expert Systems with Applications, 27(3), 2005, [10].Vanstone Bruce and Gavin Finnie, An empirical methodology for developing stock market trading systems using artificial neural networks, Expert Systems with Applications: An International Journal Volume 36, Issue 3 (April 2009) [11].B. LeBaron. Empirical regularities from interacting longand short-memory investors in an agent based stock market. IEEE Transactions on Evolutionary Computation, 5(5): , October [12].B. LeBaron. Evolution and time horizons in an agent based stock market. Macroeconomic Dynamics, [13.B. LeBaron. A builder s guide to agent based financial markets. Quantitative Finance, 1(2): ,2001. [14].Harry Markowitz, Portfolio Selection, Journal of Finance, March 1952, pp Prof PN Kumar 1, MTech(IITK),FIE,FIETE is Chairman CSE Dept, Amrita School of Engineering, Ettimadai (Amrita Vishwa Vidyapeetham). He is a visiting professor at Amrita School of Business and is an adjunct faculty of State University of New York, Buffalo. Interest areas are Agent Based Modeling, Soft Computing and Software Project Management. Rahul Seshadri.G 2 is a Graduate in Computer Science & Engg from Amrita School of Engineering, Ettimadai (Amrita Vishwa Vidyapeetham). Interest areas are Stock Market Prediction and Soft Computing. Presently he is pursuing his MBA. Hariharan.A 3. is a Graduate in Computer Science & Engg from Amrita School of Engineering, Ettimadai (Amrita Vishwa Vidyapeetham). Interest areas are Stock Market Prediction and Software Engineering. Presently he is a Software Developer in Cognizant Technologies at Coimbatore. Dr VP Mohandas 4,, PhD(IITB),FIETE is Chairman ECE Dept, Amrita School of Engineering, Ettimadai (Amrita Vishwa Vidyapeetham). His areas of interests include Dynamic System Theory, Signal Processing, Soft Computing and their application to socio-techno-economic and financial systems. Dr P Balasubramanian 5,Ph.D (JNU) is Associate Professor, Finance at Amrita School of Business. He has thirteen years of experience in teaching and research and two years of industrial experience. Interest areas are Micro Economics, Macro Economics, Options, Futures and other Derivatives.

336 314 Use of XML to Persist and Transfer Offline Data For Personalized Profiles Sandeep Chawla 1, Suchita Goyal 2 1 B.E. Information Technology, University of Delhi Working as Software Engineer, Itaas Inc. 2 B.E. Information Technology, University of Delhi Working as Software Engineer, Servigistics Inc. ABSTRACT As the internet proliferates a web of interlinked data wider across the globe and deeper into the remotest areas, many more users across the globe are getting access to the web. Wit the advent of Web 2.0 the pattern of internet usage has undergone a paradigm shift. Every user enjoys a virtual presence by creating personalized profiles at different portals like social network websites Facebook, LinkedIn etc & as in our case faculty/employees profiles at College/Organization. These profiles act as a repository of information of the user and help the employers/students to choose their prospective employees/institutions. Thus these profiles need to be regularly updated over the internet. At times, when a user wants to update his/her profile, there is a chance that net could be disconnected or very slow for images and publications etc. to upload, or there could be a chance that the user wants to have a full preview of his updates and uploads, before he actually wastes his bandwidth and hours on it. Keeping in mind these points of views we started to develop a desktop application that would aid the faculty at Delhi College of Engineering in updating their profiles at official website with ease. Through this paper we present user profile manger in the form of a desktop application. The tool is specifically engineered to facilitate the faculty members to regularly update their profiles and other auxiliary information on the college website remotely via their desktop. This tool is predominantly developed in C# programming language and relies on Microsoft.NET Framework 2.0. The application highlights the interaction of C# with a remote database server and ability of XML to persist data at client side and transfer the same also. We present a detailed look into the architecture of our application and then implement a prototype of the same. Keywords: Profile manger, C#-Database interaction, offline profile preview, offline update 1. INTRODUCTION Nowadays the World Wide Web (WWW) tends to become the major medium by which users access information and services[3]. The format of the information can be present in different ways and is dynamically changed primarily being governed by user preferences. The onset of Web 2.0 gave the freedom to a user to personalize the web. Web personalization can be defined as any action that tailors the Web experience to a particular user, or set of users[4]. This may be carried out in a number of ways and creating personalized profiles is one of the many ways to create net presence. These profiles not only help the users to interact with each other but at times are a source of valuable information. This service requires user to regularly update their profiles over the net so that proper information about them is passed on. A similar concept of profile creation and maintenance has been in use since few years in the academic field. Many colleges create the profiles of its faculty members over the internet.

337 315 This information about a professor serves to be of use to various companies and students and thus needs to be updated on a regular basis. Managing and updating a profile on Web 2.0 platform is relatively easy but poses concern when the same profile needs to be updated regularly and traffic considerations are a serious concern. It requires the user to know the web technologies with which the site was developed like PHP, HTML et al, unless provided with a user-friendly user interface. Thus it amputates a user s efficiency. Thus we developed a desktop application, which facilitates a user to update his/her profile remotely. This application provides a user (in this case a faculty member) an interface which contains the same profile as on the website. It stores the updates and data of a particular user on his/her system locally and whenever the internet connection is available for the next time all the information which includes the updates are transferred in a XML file to the server where the website is uploaded. The corresponding updated fields are then reflected on the website. Also, the tool provides the faculty with various functionalities like updating their resumes, publications, editing images and adding lecture notes. The rest of the paper is divided as follows in the next section we define the problem statement along with an analysis of proposed solutions. In section III we will elucidate the methodologies of development procedure and the various technologies that were applied during the development procedure. Followed by the last section, IV, which gives the future scope of the project and conclusion. 2. PROBLEM STATEMENT The official website of Delhi College of Engineering ( consists of a number of sections that focus on various aspects of college like academic programs being offered, departments, activities, et al. The site also has a section which contains details of faculty & staff. Every faculty member has his/her own profile on the web site. The profile page of any contains key information about his/her profile like Name and designation, Image, , Lecture notes, Publications, Books Published, Research Publications in Conferences and in International Journals etc. All this data needs to be updated often which proves utilitarian for the students scouting out for thorough guidance in some specialized discipline. To do so, the faculty sends the information to be updated to the department which forwards it to the concerned web management team and all this procedure takes time. Since long there has been a requirement for such an application that can help the faculty to do the described at their own with no inconvenience and an immediate effect. 2.1 ANALYSIS OF PROPOSED SOLUTIONS The initial solution comprised of creating a FTP account for every faculty with his/her rights restricted to his/her folder on the website. This solution stated that whenever a faculty needs to overhaul his/her profile, he/she needs to download the corresponding PHP page using FTP and then try to understand the code in order to edit. Unreasonable for any faculty member, who is unaware of the web technologies used to first learn those technologies in order to update his/her profile. Also allowing access to FTP of the college website can prove to be a major security threat & opens a door to a number of vulnerabilities. Another solution could be to develop a web application which would allow access to a form like page post user log in. There he/she can update his/her respective profile in a simple and easy manner as is in the case of social

338 316 networking website or updating an account profile. But this solution necessitates a user to be connected to the internet which is not always the case. Thus we tried to develop a desktop Application which will reside on user s computers and help them update their profiles on The application works both in online as well as offline modes. If an internet connection is not available then the user can create/update his profile which gets saved as an XML file on the user s system. He/She can then even have a preview of how these updates will be reflected on the website. Once the internet connection is available, the C# functions parse the XML file to get the desired information which is then passed to the server to update the respective PHP pages. The various services provided are enlisted below and a detailed explanation on each is given in the Section IV. Faculty members can create a new profile online as well as offline. Update the profile online as well as offline. Edit the image for his/her own page at the website. Manage his/her own publications at the website. Upload their lecture notes on respective profile page for the perusal of others. Send via <FacultyMemberName>@dce.ac.in account. 3. TECHNICAL ASPECTS & APPLICATION DEVELOPMENT 3.1 TECHNICAL ASPECTS Our work has the following steps which taken together constitute the technical contributions of this paper. From the vast pool of programming languages C# was chosen for application development as it surpasses other programming languages in terms of enhanced features, like faster and userfriendly UI development, more efficient than other standard languages like java and runs faster than those too. C# supports the introduction of XML comments. Far from being just another way to add comments to code, XML comments can actually turn into your documentation. The comments are placed into XML format and can then be used as needed to document your code. This documentation can include example code, parameters, and references to other topics. It finally makes sense for a developer to document his or her code, because those comments can actually become documentation independent of the source code.[12] STORING DATA ON LOCAL HOST The first and foremost problem faced was storage of profile data on client s computer. Installing any database server for the same didn t seem to be a good choice because it requires fair amount of memory and continual querying of database for data display and updating. Moreover, there was no standard means of transportation of database data to a web server running at remote host. Serialization of data can be looked upon as a tentative solution but considering the fact that the language which will serialize it at client s end is C# and the web scripting language used in PHP, the suggested solution fails to make an impact. After database server, the next obvious choice was use of extensible Markup Language (XML). Today XML has become the unquestionable standard for generically data to be shared. XML provides a great way to take a snapshot of the object and store the contents into persist-able form a file or a database typically. Also, C# provides wide support for the use of XML in an open, standards-compliant manner.

339 317 These extensive features of XML have been utilized in our application to provide support for management & use of data on remote servers when online, and store & use data while offline. The profile information is stored on the remote server in the form of multiple tables and on user s machine, in XML format. XML tags are defined in accordance with the information to be displayed. Use of XML made it possible to combine multiple tables into a single document. In our application, a XML file is created locally when user logs in for the first time. It has user authentication information (encrypted) besides profile details. Once all the details are fetched, user can view them over application s User- Interface. To present this xml data in the format as on the web, our application harnesses the power of extensible Stylesheet Language Transformations (XSLT). It has become the language of choice for a very wide range of XML applications. XSLT is a style sheet language for XML documents. It can be used to process multiple XML documents and to produce any combination of text, HTML and XML output; XML is transformed on the fly without the user even noticing. XSLT uses a template-driven approach to transformations: you write a template that shows what happens to any given input element. [11]. With XSLT we have added as well as removed elements and attributes to or from the output file. We have also rearranged and sorted elements, performed tests and made decisions about which elements to hide and display, etc. In the transformation process, we created an XSL Style Sheet with a transformation template and added the XSL style sheet reference to the XML document. Since the components provided by C# are XSLT compliant so it nicely transformed our XML into XHTML. It helped us add, delete or edit the information using the same interface; the changes done are saved into the XML along with a flag value indicating the type of action, insertion/deletion/update/no effect. When user wants to persist the changes, dataset is created to point to the view in database; it is updated using the XML stored locally. After updating the dataset, the changes are persisted into the database. All database activities are done atomically in separate transaction to ensure consistency of the database. 3.2 APPLICATION DEVELOPMENT When user runs the application for the first time he is asked to authenticate himself on the remote server. If the user successfully logs-in, then his profile data is fetched from database and stored on his computer in XML files. And the user is directed to ProfEdit s main window as shown in figure 3. The window has following six tabs: Fig. 3 Application s Main Window. Fig 4 XSLT of the application under consideration

340 318 Fig 5 XML of the application under consideration 1) Profile Preview: This tab shows the profile in the same format as on the college s website. The data displayed is fetched from XML files and is formatted in the desired format using XSLT. 2) Basic Update: This tab shows the minimum but mandatory fields for the existence of profile on the website. User can fill the fields provided against the desired titles and can update locally to preview the changes or can directly update on server. For updating on server, user needs to authenticate himself. An undo option is provided to undone the changes done during the session. Session starts when user authenticates himself for the first time and ends when user exits the application i.e. once the user has logged in from any of the tab, further updating in the same tab and the others, can be done by simply clicking on the button. This is because every tab has similar updating options. This session state can be maintained by creating a file similar to cookie file on client s machine. 3) Maintaining Publications: This tab helps user to add new publication(s), delete or edit the existing ones and undo the changes. Updating can be done locally or directly on server as user wishes. 4) Editing Image: This tab provides image browsing and editing (image cropping etc) options. Almost all image formats are allowed to be browsed. After edting, the image is uploaded directly to FTP server. 5) Managing Lecture Notes: This tab helps faculty to manage lecture notes. All file formats are supported. The options provided are adding new one(s) and deleting the existing ones. There is no limit on the number of notes. 6) DC This interface makes user to log on to the SMTP server by providing his username & password for their respective accounts at After authentication, the user can generate the mail message using the form provided having the interface similar to that provided by the mail server. 4. FUTURE SCOPE AND CONCLUSION 4.1 FUTURE SCOPE The generation, encryption and transfer of data over the network can be optimized by the use of multithreading paradigm. Different threads can be created which are monitoring each of these tasks individually. We can accommodate for live update feature as and when a user updates the profile in desktop application, rather than just clicking a submit button after he has finished. We can use AJAX for this live synchronization with the webservers provided the user is online. The encryption and decryption algorithms can be modified and made more secure against vulnerabilities. The application currently doesn t have any formal security mechanisms except for those offered by the basic authentication. 4.2 CONCLUSION XML has always been considered the best means for inter-language communication, but harnessing its efficacy to persist data can open an arena for plethora of applications. Applications can also talk to each other with the use of XML, regardless of the fact that one application is web-based and other is desktop

341 319 based. Thus XML opens doors for communication across languages and applications which are not just on different machines but different Operating Systems also. 5. REFERENCES [1] Pederson, J. H., Watson, K., Nagel, C., Skinner, M., White, E. and Reid, J. D., 2006, Beginning Visual C# 2005, Wiley Publishing Inc., USA, , , Part three and Part five. [2] Robinson, S., Watson, K., Nagel, C., Skinner, M., Glynn, J., and Evjen, B., 2006, Professional C#, Wiley Publishing Inc., USA, Part two. [3] Yu, Jie; Luo, Xiangfeng; Xu, Zheng; Liu, Fangfang; Li, Xuhui; Representation and Evolution of User Profile in Web Activity, IEEE International Workshop on Semantic Computing and Systems. [4] Bamshad Mobasher, Honghua Dai, Tao Luo, Yuqing Sun. Integrating Web Usage and Content Mining for More Effective Personalization. Proceedings of the First International Conference on Electronic Commerce and Web Technologies Pages: [5] Welling, L. and Thomson, L., 2001, PHP and MySQL Web Development, Sams Publishing Inc., USA, Part I and Part - II. [6] Patrice, P., 2006, Microsoft Visual C Sharp 2005 Build a Program Now, Microsoft Press, USA, Part-I. [7] Deitel, H. M. and Deitel, P. J. - Deitel & Associates, Inc., 2005, Visual C# 2005: How to Program, Second Edition, Prentice Hall Publication USA. [8] Kay, M., 2000, XSLT Programmer's Reference, Wrox Press Ltd., USA, chapter 4. [9] Williams, M., 2002, Microsoft Visual C#.NET, Microsoft Press, USA, Part IV. [10] Rustan, K. and Leino, M., 2005, Exception safety for C#, Microsoft Research, USA. [11] [12] Sandeep Chawla is Bachelor of Engineering (B.E.) in Information Technology from Delhi College of Engineering, University of Delhi, Batch of A Software Developer by profession, he likes to work on latest technologies and emerging trends. He has successfully completed projects at organizations like Servigistics, Slideshare, Ministry of Environments and Forests (Govt. of India) and PricewaterhouseCoopers. Currently, he is associated with Itaas Inc. Suchita Goyal is Bachelor of Engineering (B.E.) in Information Technology from Delhi College of Engineering, University of Delhi, Batch of She has worked on and is wellproficient with a number of technologies and frameworks. She has successfully completed projects at various organizations like Ministry of communications and information technology, Planning Commission (Govt. of India), Delhi College of Engineering and CMC Ltd. Currently, she is associated with Servigistics Inc.

342 Role of Knowledge Management in Enhancing Information Security 320 Yogesh Kumar Mittal 1, Dr Santanu Roy 2 and Dr. Manu Saxena 3 1 Ajay Kumar Garg Engineering College Ghaziabad,Uttar Pradesh, India Research Scholar, Singhania University Jhunjhunu,Rajasthan,India 2 Institute of Management Technology Ghaziabad,Uttar Pradesh, India 3 Human Resource Development Centre Ghaziabad,Uttar Pradesh, India Abstract User s knowledge of information security is one of the important factor in information security management as 70-80% security incidents occurred due to negligence or unawareness of users. In this paper we have analyzed the utility of knowledge management tools to rapidly capture, store, share and disseminate the information security related knowledge with the view that it should be effectively applied by the information system users. We found that the knowledge management tool can be used to enhance the information security. Keywords: Knowledge Management, Information Security, Knowledge Management Tools, Information Security Challenges. 1.0 Introduction Due to fast pace of change in IT technology and its important applications, new security threats evolves around it. New and smart methods of information security are also devised by researchers to mitigate the risk occurred due to these threats. In the last decade process based information security management system(isms) such as ISO27001 and COBIT have emerged. Many organizations since then have adopted such ISMS. Knowledge Management(KM) is another management discipline enterprises employ, with aim to foster a more effective management of knowledge[1]. Organizations sometimes spend substantially on firewall, proxy, antivirus, intrusion detection mechanism, digital signatures, special network devices and protocols etc., assuming that security of information can somehow be ensured by procuring these technology solutions from the market. This is a wrong notion because security management is more of managing an end-to-end system rather than just installing technical solutions. As like any other full-fledged system, this has many components including people, policies, procedures, processes, standards and technology[2]. Information may be stored in a server, PC, Laptop, mobile phone or in any other device, it may be in transit from one place to another place through some communication channel, or may be under processing through a program, security of the information may be breached at any stage. Confidentiality, integrity and availability are the three major information security considerations. Protection of information is just not dependent on only information security people of the company but all the users. All the user of information system are like on the gates of a building and the gate opening by any of the employee may prove fatal for the safety of the whole information system. Inspect the domestic and foreign each type of information security event to discover that 70-80% are because the internal personnel negligence or intends to divulge creates, 20-30% are because the hacker invades or other external reasons creates[3]. 2.0 Information Security Challenges In general, the information security management of an organization broadly deals with the processes and procedures that the employee should adhere to in order to protect the confidentiality, integrity and availability of information and other valuable assets. The standard

343 321 approach to managing information security involves conducting a risk analysis to identify risks to confidentiality, integrity, and availability of information systems, which is followed by risk management where safeguards are employed to mitigate those risks[4].with this definition in mind, the main goals of information security within organizations are to reduce the risk of systems and organizations ceasing operations; maintain information confidentiality; ensure the integrity and reliability of data resources; ensure the uninterrupted availability of data resources and online operations; and ensure compliance with national security laws and privacy policies and laws. Thus, management of information security involves implementing and maintaining information security policies and procedures to minimize opportunities for threats like computer fraud [12]. Knowledge of information security is essential for all the employees or users as per their requirements. Lot of information on information security is available through books, internet, journals etc., but people don t use this information because: 1. Getting particular useful information out of a glut of information is very difficult and time consuming. 2. Users may not aware about importance of information security, they feel that it is the work of information security staff or IT department. 3. Lack of motivation in getting Information security knowledge. 4. Information Security experts not willing to share the knowledge with the users. 5. Communication gap or social gap between users and experts. 6. Users may not know, who is the expert for particular security issue. 7. Users are very busy in their regular work. 8. Infrastructure not available to communicate. 9. Geographical distance between users and experts. 10. Lot of knowledge is experience based or in tacit form and need to be codified to be shared or require a proper platform to share. So, there may be number of reasons of unawareness about security threats and remedies. But the result is really horrible in terms of security incidents which may lead to information leaks, non availability, compromise on integrity etc. and huge losses in terms of reputation, loss of customers and direct monitory loss[5]. 3.0 Role of KM in Information Security The role of Knowledge management is really important to manage the knowledge of information security as Knowledge Management has been defined as the capability by which communities capture the knowledge that is critical to their success, constantly improve it, and make it available in the most effective manner to those who need it [13]. According to Granneman [6] most people do not secure their computers or act in a secure manner. The main reason being that the average user just does not know what to do. This is alarming, considering that 65.9% of the Australian population are Internet users [7] and the success of the 2000 LOVELETTER virus and 2003 Blaster/SoBig worms were largely due to individuals uneducated in information security issues [8]. A holistic understanding of e-security and privacy issues is vital for the individual as well as for the society. Therefore key considerations and possible solutions include: Education and Awareness Education and awareness efforts targeting existing and emerging new threats, risks, vulnerabilities, countermeasures and safeguards are required. Foster a security conscious culture A security culture where compliant attitudes, behaviors and sensitivity to privacy and security become second nature and assumed throughout every day life[9]. Knowledge management is enabling and enhancing capabilities to perform such processes, including sourcing and deployment of the right knowledge assets, in order to achieve the desired results. Knowledge assets include embodied knowledge in people; embedded knowledge in technology, systems and processes; enculturated knowledge in work relationships, teams and networks; and actionable information and insights[10]. Three major phases of KM cycles are: 1. Knowledge capture and /or creations 2. Knowledge sharing and dissemination 3. Knowledge acquisition and application Knowledge capture refers to the identification and subsequent codification of existing knowledge and know how with in organization and/or from the environment. Knowledge creation is the development of the new knowledge and know how or innovations that did not have previous existence with in the company like from experts, research papers etc.. Once it has been decided that the new or newly identified knowledge is of sufficient value, the next step is to contextualize this content. Contextualize means giving link to the contributors of that knowledge and tailoring it towards the target users. Then the

344 322 knowledge is shared and disseminated. The knowledge is disseminated to the users using portals, s and other KM tools. Users apply the knowledge and with their experience also new knowledge is generated and captured. Knowledge management advocates different type of recognitions and incentives for the people who are sharing their knowledge and the people who are actively using the knowledge to improve their knowledge and performance. Knowledge management has multi dimensional benefits at different levels from individuals to organization such as for individuals, it helps people to do their job in an efficient way through better decision making and problem solving. Help people to keep up to date. On community level, promotes peer to peer knowledge sharing. At the organization level, diffuse best practices, builds organizational memory. In other words, knowledge of how an organization functions in the context of management of information security can significantly impact the effectiveness of procedures in minimizing opportunities for computer fraud. 4.0 Application of KM Tools for Information Security: In order to design successful tools for knowledge sharing, a strategy needs to be chosen. Hansen et al. distinguish two main knowledge management strategies: codification and personalization. Codification is the people-to documents strategy. Here the effort is to load intranets and databases with best practices, case studies and how-to guides to help people in their day-to-day work[11]. Personalization is the people-to-people strategy. Here the effort is to link people with other people and to grow networks and community of practices. Emphasis is on informal-knowledge sharing. Following KM tools may be used for improving information security: 1. Content Management 2. Knowledge Taxonomies 3. Groupware 4. Online Communities of Practice 5. Enterprise Portal 6. Social Network Analysis and Design 7. E-learning 8. Storytelling and Narrations 9. Wireless tools for knowledge Mobilization 10. Innovation and idea management system 11. Tools for extending KM across organizational boundaries 4.1 Content Management : A well designed content platform must be able to handle multiple content types, sources and access patterns. Theses content sources include security related libraries, activities and personnel directories. Content can be structured or unstructured. Some of it is generated online during various knowledge activities(e.g. on line brain storming). Organizations may use content management system for information security best practices, lesson learned, security case studies etc.. Content teams, meta data, knowledge maps, and a workflow contextualization can ensure effective reuse of the content. Advanced content management system include features for seamless exploration, authoring templates, maintaining integrity of web pages and links, periodic review, archiving, metadata, version control, rule setting, indexing, audits, authorized access, administration alerts, and flexible repurposing for multiple platforms and formats. 4.2 Knowledge Taxonomies: Taxonomy is the practice and science of classification according to natural relationships. The info-glut or digital sprawl on corporate intranets has led to users not being able to find relevant information in time and numerous taxonomy development tools are coming to the rescue. It must reflect the needs, behaviors, tasks and vocabulary of the users, and be able to provide multiple paths and points of view. Taxonomy should be easy to maintain and users should find it easy to understand, navigate and contribute. It will help the users to easily locate specific information security knowledge. 4.3 Groupware: Desirable features for collaboration in the context of KM include affinity building, knowledge mapping, threading, polling, group document creation, rating, anonymity and access management. A notable trend in tools for collaboration between networked employees is the convergence between asynchronous (e.g. collaborative document management) and synchronous(e.g. instant messaging) service. It is an important tool for knowledge sharing among the peer groups. It is an important tool to disseminate information security information instantly to a group like information regarding new virus attacks. 4.4 Online Communities of Practice(CoP): Online communities constitute a growing part of the organizational landscape of 21 st century global players, but

345 323 businesses are still at the early stages of individual and organizational optimization of web based communities. Online CoPs are emerging as powerful tool for knowledge exchange and retention. Participation levels in CoPs can be segmented into core, active, and peripheral. Success levels can be diagnosed via the application of knowledge., in the form of interviews anecdotes and employee survey. Expertise directories are a useful way for connecting knowledge worker in such forming communities, but they must connect people and not just resumes. CoPs are particularly useful in discussing current security related problems and come out with solutions. 4.5 Enterpriser Portal Portals help create the on demand workplace, customized to individual employee needs. A well-designed portal can serve as a delivery channel for KM applications any time, any place, and on any device. Knowledge portals are the single point of interaction and coordination for collaboration. General user may reach the portal for getting their solutions of security related problems and current security scenario. 4.6 Social Network Analysis and Design Social network analysis (SNA) is emerging as a powerful tool for mapping knowledge flows and identifying gaps. SNA can be used to reinforce existing flows and to improve knowledge integration after activities like mergers and acquisitions. Natural language techniques, visualization tools, and recommender systems can be harnessed here, leading to actions like identifying key individuals for retention or expended roles or creating teams for cross-organizational and cross-functional activities. Direct applications of SNA include security process redesign, role development, and improved collaboration between knowledge seekers and providers. SNA can help identify central people, connectivity levels of individual knowledge workers, diversity of subgroups, and level of organizational inter-connectivity. Getting things done often depends less on formal structure than on informal net-works of people. SNA can help improve general security environment by disseminating information security knowledge naturally and effortlessly. 4.7 E-learning One interesting emerging development on the KM front is the growing convergence of viewpoints between the KM community and the e-learning community. The concept of KM can be united with the goals of e-learning to create the larger ideal of a learning organization-via blended learning, skills directories integrated with course delivery, and the interleaving of working and learning. KM and learning management are two complementary disciplines that are continuously growing closer and support an innovative and agile enterprise. For training of new recruits about information security and for training of new security technologies, e-learning may be very useful. 4.8 Storytelling Narratives Personal storytelling builds community and can revitalize the way we do business. Non-traditional business communication techniques like art, theatrical tools and even a poetry can improve internal and interpersonal communication. Stories are good framework for sharing information, meaning and knowledge. Blogs encourage story-telling and foster understanding because they usually offer context. Social engineering type of attacks can be easily described using these techniques. 4.9 Wireless Tools for Knowledge Mobilization One of the most notable emerging trends in workforce connectivity is the increasing use of mobile technologies to take KM to another dimension- knowledge mobilization -by bringing relevant knowledge directly to the fingertips of a company s road warriors and fieldworkers via cell phones, PDAs, industry-specific handheld devices, Wireless Local Area Network (WLAN), and Radio Frequency Identification (RFID) tags. While personal computers (PCs) and workstations have come under some criticism for tethering knowledge workers to their desks, wireless technologies may be the perfect answer to mobilizing the workforce by letting them capture and harness key information and knowledge attributes wherever they are, whenever they want, and however they want. This tool enables information security knowledge to be disseminated for the people on the move and it is immediate Innovation and Idea Management Systems Managing an innovation pipeline, promoting an idea central or ideas marketplace, and creating the hundred headed brain are some creative approaches being adopted by KM pioneers. KM also helps organizations increase the efficiency of innovation by improving access to experts and tapping into past innovations. New innovative ideas and information security solutions can be evolved using these systems Tools for Extending KM across Organizational Boundaries

346 324 Online services such as dial-up bulletin boards and web communities have actually helped network communities of interest across the globe for years. The world Bank has leveraged a strategy of global knowledge, local adaptation for brokering global knowledge exchanges. Information Security knowledge can be accessed from all over world to be applied in the company. 5.0 Conclusion: We can see that to deal with the ever changing nature of information technology and the newer security threats coming up at a very fast pace, we need some technique to educate the users in an effective manner. KM tools can be used to evolve newer, economical and faster methods to deal with information security issues. KM tools like content management may be used to create content and update information security knowledge like information security standards and best practices, taxonomies to easily understand and locate the right and required information, CoPs for consulting with each other and giving a feeling of belongingness to share the knowledge. Enterprise portals can be used as a single point of contact for all the interested stakeholders. E-learning methods may be used to educate the new joiners and to train on the latest developments in the area. Storytelling is good for understanding point of view and social aspects. Wireless tools make the person free from a specific location and person on the move may get the latest knowledge. Innovation is the key for the new solutions. Lot of research and innovations are taking place in information security field. KM can encourage people to give new ideas and rewarding them accordingly. This way we can see that there is a lot of scope to improve information security using knowledge management techniques. References: [1]Knowledge-Centric Information Security, Walter S. L. Fung, Richard Y. K. Fung, 2008 International Conference on Security Technology, IEEE [2]Information Security Management - A Practical Approach,2007,Manik Dey,, IEEE [3]Behavioral science-based information security research, Yang yue jiang Yu yong xia, 2009, First International Workshop on Education Technology and Computer Science IEEE [4]Knowledge Based Model for Holistic Information Security Risk Analysis,2008 Wen Huang, Yong-Sheng Ding, Zhi-Hua Hu, Jing-Wen Huang, 2008 International Symposium on Computer Science and Computational Technology, IEEE [5]Knowledge management within information security: the case of Barings Bank, Shalini Kesar, International Journal of Business Information Systems Vol. No.6 pp [6] S. Granneman, "A Home User's Security Checklist for Windows,"SecurityFocus, [7]Nielsen/NetRating, "Top Rankings," Netrating, Inc., [8]CERT/CC and Carnegie Mellon University, "CERT/CC Overview Incident and Vulnerability Trends," [9]The Multifaceted and Ever-Changing Directions of Information Security Australia Get Ready!, Leanne Ngo and Wanlei Zhou, 2005 Proceedings of the Third International Conference on Information Technology and Applications (ICITA 05) IEEE [10]Knowledge Management in Asia: Experience and Lessons 2008,Report of the APO Survey on the Status of Knowledge Management in Member Countries. [11] Collaboration and Knowledge Sharing Platform for supporting a Risk Management Network of Practice Katerina Papadaki, Despina Polemi, 2008, The Third International Conference on Internet and Web Applications and Services, IEEE [12] BSI 2002 [13] Birkenkrahe, M. (2002). How large multi-nationals manage their knowledge. Business Review, 4(2), pp Yogesh Kumar Mittal did B.Tech. from Maulana Ajad College of Technology, Bhopal, India (Now MANIT, Bhopal) in 1987 than M.Tech. in Computer Science and Technology from University of Roorkee, Roorkee, India (Now IIT Roorkee) in He also did PGDBM from IMT, Ghaziabad, India in He qualified prestigious CISA (Certified Information System Auditor) exam in He has around 21 years of experience in industry and academia. He has worked as Consultant, Information System Auditor, General Manager and Chief Executive Officer before joining the teaching profession. He published 10 papers in National/International conferences/journals. His academic and research interest includes IT in Business, Knowledge management, Software Project Management, Enterprise Resource Planning, Software Engineering, Information security and Auditing, Social and Cultural issues. Dr. Santanu Roy is currently serving as a Professor, Operations Management Area, at Institute of Management Technology (IMT), Ghaziabad, India. Dr. Roy had earlier served as a Senior Scientist (Scientist F) in National Institute of Science, Technology and Development Studies (NISTADS), New Delhi. Dr. Santanu Roy has done his Ph.D. in Industrial Engineering and Management from IIT Kharagpur, India and Integrated Master of Science (M.S.) from IIT Delhi. He has more than 26 years of experience in research, consultancy and teaching. Dr. Manu Saxena did B. Sc. in 1977 from, Meerut University, India, M. Sc. in 1979 from University of Roorkee, Roorkee, India Ph. D. from University of Roorkee, Roorkee, India in Operational Research in He published 19 papers in national/international conferences and journals. He supervised 13 dissertations of post graduation level.

347 ISSN Online: Cloud Computing for Managing Apparel and Garment Supply Chains - an Empirical study of Implementation Frame Work Prof. A.K. Damodaram, Department of Mechanical Engineering Sree Vidyanikethan Engineering College, A. Rangampet. Chittoor (Dt) A.P. Mobile Prof. K. Ravindranath, Department of Mechanical Engineering, S.V. University College of Engineering, SriVenkateswara niversity,tirupati a.p. Abstract With operating efficiencies already quite high, members of the apparel and manufacturing supply chain need to look at strategies other than reducing labor costs to improve margins. The key area is collaboration with supply chain partners. Collaboration offers an opportunity to reduce costs in the supply chain in the areas of product development, inventory holding, and manufacturing through better capacity utilization, lower reject rates, fewer chargebacks and profitability. Apparel and garment industry is highly fickle and are characterized by short life cycles, unpredictable demand, whimsical consumers, multiple trading partners, difficulties in doing cross-border trade and stagnating economic conditions. Information Technology enabled collaboration should be the key area for tackling the hurdles in this regard. Cloud computing as an IT enabled option for model for delivering ondemand, self-service computing resources with ubiquitous network access, locationindependent resource pooling, and rapid elasticity. In order to make the supply chain efficient, collaboration among partners is becoming a necessity. Retailers should prefer to form partnerships with suppliers who have gone ahead and implemented processes and systems which facilitate collaboration. In this paper, an empirical study of adaptability of Cloud computing model for apparel and garment manufacturers to achieve collaboration among the supply chain partners to manage the Supply Chain is presented. Key words: Cloud Computing, SCM, Collaboration, cloud computing framework. 1. Cloud Computing: Definitions of Cloud computing on the Web: Definition 1. Cloud computing is Internetbased computing, whereby shared resources, software and information are provided to computers and other devices on-demand, like electricity[1] Definition2. Computing in which services and storage are provided over the Internet [1] Definition3. On-demand self-service Internet infrastructure where you pay-as-you-go and use only what you need, all managed by a browser, application or API. Cloud computing is broken up into multiple segments including: Cloud Infrastructure, Cloud Platforms and Cloud Applications. [2] Definition4. A new generation of computing that utilizes distant servers for data storage and management, allowing the device to use smaller and more efficient chips that consume less energy than standard computers. [3]

348 ISSN Online: It is a style of computing where massively scalable IT-related capabilities are provided as a service using Internet technologies to multiple customers. Cloud Computing, a key differentiating element of a successful information technology (IT) is its ability to become a true, valuable, and economical contributor. It implies a service oriented architecture, reduced information technology overhead for the end-user, greater flexibility, reduced total cost of ownership, on- demand services and many other things. While there is no arguing about the staying power of the cloud model and the benefits it can bring to any organization, mainstream adoption depends on several key variables falling into alignment that will provide users the reliability, desired outcomes, and levels of trust necessary to truly usher in a cloud revolution. Until recently, early adopters of cloud computing in the public and private sectors were the catalyst for helping drive technological innovation and increased adoption of cloud-based strategies, moving us closer to this inevitable reality. Today, driven in large part by the financial crisis gripping the global economy, more and more organizations are turning toward cloud computing as a lowcost means of delivering quick-time-to-market solutions for mission-critical operations and services. The benefits of cloud computing are hard to dispute: and electronic storage resources, as services over the network. As a Platform as a Service (PaaS), includes tools and environments to build and operate cloud applications and services; As a Software as a Service (SaaS), enables on-demand use of software over the internet and private networks; and As a Business as a Service (BaaS), includes application functionality coupled with physical and human resources required to perform a broader set of business activities typically a major module of activity in a roader business process (e.g., a call center module, as part of the customer service process), or in some cases the complete business process itself (e.g., fully cloud-based supply chain management). These models of computing are being driven by the confluence of several changes in the business environment and IT landscape. From the business perspective, the trend towards consumer-driven innovation and partnership ecosystems is accelerating software development timeframes. Simultaneously, from the IT perspective, several trends focused on increasing the efficiency of software distribution and hardware utilization have converged to enable a cloud computing model, notably early adoption of Software as a Service, proliferation of Hardware Virtualization, and the advent of Utility Computing. 1. Reduced implementation and maintenance costs 2. Increased mobility for a global workforce 3. Flexible and scalable infrastructures 4. Quick time to market 5. IT department transformation (focus on innovation vs. maintenance and implementation) 6. Greening of the data center 7. Increased availability of highperformance applications to small/mediumsized businesses Rapid experimentation by early cloud providers has created four distinct layers of services: Infrastructure As a Service provider (IaaS), Includes raw utilities such as compute power Cloud Computing is more than a technology. It is more than a platform. It is more than just a hosting provider. It is more than just an application hosted as a service. It is more than providing storage services on the Internet. It is a combination of all the above. In today s economic environment as Enterprises try to balance out and optimize their IT budgets, Cloud computing can be an effective strategy to reduce the IT operations and management costs and free up critical resources and budget for discretionary innovative projects. Typically, Enterprises have a 80/20 split between regular ongoing IT operations cost which includes hardware, software licensing costs, development, data center maintenance etc Vs new investment for solving critical business needs which is critical for businesses to survive in these challenging times. Cloud Computing can have a significant

349 ISSN Online: impact in this by reducing the footprint of IT operations by taking out the upfront capital investments needed for hardware and software licensing. It enables a Use what you Need and Pay for what you Use cost model. This will enable businesses to invest on innovative solutions that will help them address key customer challenges instead of worrying about operational details. Even though IT budgets are being slashed, enterprises cannot afford to stop investing in IT because IT is what helps them gain and maintain a competitive advantage. The Cloud offerings will help enterprises to continue to invest in IT without having to take up big budget and long term IT projects. Investment in IT changes from being a Capital Expenditure to Operating Expenditure. Enterprises can become agile and harness the power of Information Technology to drive unprecedented customer value. Fig.1. Cloud Computing model. 2. What technologies constitute cloud computing? Cloud Computing is a paradigm that is composed several Strata of Services. These include services like Infrastructure as a Service, Storage as a Service, Platform as a Service and Software as a Service. Different Cloud Providers have developed various access models to these services. The access to these Services are based on standard Internet Protocols like HTTP, SOAP, REST, XML and the infrastructure is based on widely used technologies including Virtualization, hosting. Cloud Computing is the maturation and coming together of several prior computing concepts like Grid Computing, ASP, Server Hosting, Utility Computing and Virtualization. 3. Supply chain of Apparel / Clothing/ Garment sector The textiles and clothing sectors can be seen as a supply chain consisting of a number of discrete activities. Increasingly the supply chain from sourcing of raw materials via design and production to distribution and marketing is being organized as an integrated production network where the production is sliced into specialized activities and each activity is located where it can contribute the most to the value of the end product. When the location decision of each activity is being made, costs, quality, reliability of delivery, access to quality inputs and transport and transaction costs are important variables. Apparel and Garment manufacturing industry environment is characterized by: Entire demand for a given season/style order must be fulfilled by one lot

350 ISSN Online: Demand and pricing for seasonal and/or fashion items is uncertain and time-sensitive Multi-tier, disaggregated suppliers, due to specialization/cost structures/quota constraints, erode loyalty and make supply chains fragile and dynamic Moderately high set up or changeover costs, so cost is lot size dependent Distributed, global suppliers; most with little infrastructure Plans based on rules of thumb for production and transit lead times, cumulative rollups/padding very inaccurate low confidence in ontime delivery creates buy and sell side date padding Delays in determining status and no easy mechanism to notify downstream suppliers or participants Lack of documented accountability and genealogy of communications Externally imposed finite capacity, e.g. time-dependent quotas, with no visibility over other competitor s shipments for same quota category Supply chain cycle time is 2-3X total season cycle times and 6-9X profit season cycle times for many styles. Consequently, consumer demand for popular styles is frequently not satisfied and margin opportunities are lost flows for product returns, servicing, and recycling; Information flows, which represent order transmission and order tracking, and which coordinate the physical flows; and Financial flows, which represent credit terms, payment schedules, and consignment and title ownership arrangements. These flows are supported by three pillars: Processes, which embed the firm s capabilities in logistics, new product development, and knowledge management; Organizational structures, which encompass a range of relationships from total vertical integration to networked companies as well as management approaches, and performance measurement and reward schemes; and Enabling technologies, which include both process and information technologies for every organization s supply chain. However, in service organizations there is usually no flow of materials, but rather a flow of documents containing the valuable information for decision making. Regardless of the type of organization, it is necessary to coordinate all the above flows among all parties involved in the supply chain using appropriate and adequate 4. Apparel/Manufacturing industry - Adaptability to Cloud Computing: 4.1 Supply Chain Management; Supply chain management is collaborative process and project management to meet the needs of the end customer efficiently and effectively. One of the key requirements of successful management of supply chains is Collaboration. It should be noted, in general, that there are three flows that occur in supply chains: Material flows, which represent physical product flows from suppliers to customers as well as the reverse 4.2 SCM in Textile/Apparel/Garment Industry: The supply chain in the textile and clothing sector is illustrated by Figure 1. The dotted lines represent the flow of information, while the solid lines represent the flow of goods. The direction of the arrows indicates a demand-pull-driven system. The information flow starts with the customer and forms the basis of what is being produced and when. It is also worth noticing that information flows directly from the retailers to the textile plants in many cases. The textile sector produces for the clothing sector and for household use. In the former case there is direct communication between retailers and textile mills when decisions are made on patterns, colors and material. In the second case textile mills often

351 ISSN Online: deliver household appliances directly to the retailers. The Indian Textile industry adds 14% to the industrial production and 8% to the GDP of India. It provides employment to 38 million people and thus, is the second largest employment provider after agriculture. The Indian Apparel & Textile Industry is one of the largest sources of foreign exchange flow into the country with the apparel exports accounting for almost 21% of the total exports of the country. Fig2. Apparel/Garment manufacturing Supply Chain. 5. Discussion on Adaptability: The clothing sector is both a laborintensive, low wage industry and a dynamic, innovative sector, depending on which market segments one focuses upon. In the high-quality fashion market, the industry is characterized by modern technology, relatively skilled designers and a high degree of flexibility. The competitive advantage of firms in this market segment is related to the ability to produce designs that capture tastes and preferences, and even better influence such tastes and preferences in addition to cost effectiveness. The core functions of firms servicing this market segment are largely located in developed countries and often in limited geographical areas or clusters within these countries. In the low to middle priced market, the role of the retailer has become increasingly prominent in the organization of the supply chain. The retail market has become more concentrated, leaving more market power to multinational retailers. These have market power not only in the consumer market, but perhaps more importantly they have considerable buying power. In addition, highvolume discount chains have developed their own brands and source their clothing directly from the suppliers, whether foreign or local. 6. Typical Cloud Platform Provider s Services and offerings: 1. Hosting 2. Storage 3. Platform 4. Application Services 5. Tools

352 ISSN Online: Table1. Key Cloud Platform Providers. 7. Adaption of Cloud Computing - Does this mean the end of IT departments in Apparel and Garment Enterprises? Cloud Computing won t make IT redundant; IT won t be defined as the way we know it today. Instead of the CIO's core focus being the infrastructure and keeping it running, Cloud will truly empower the CIO to focus on Information management. IT departments will have to focus on developing solutions and supporting business functions rapid use information to react and develop new offerings, instead of managing servers and infrastructure. Information Technologists Enterprises will need to build architecture roadmaps and develop information strategy that IT can step up to handle. In essence, IT as we know will change to more business focused than being infrastructure focused. Cloud Computing will also allow CIOs to focus on translating the business metrics endto-end and map that to technology metrics - meaning translating business goals into application and architecture goals. CIOs can have revenue goals instead of cost management goals. ROI of the application portfolio will become a key metric, and this is an area in which Cloud Computing will have a direct and pronounced impact. 7.1 What is a Cloud Maturity Model? Cloud Computing is in its infancy today and it will take 5-10 years for this to be a reality - meaning, significant adoption in the Enterprises even for critical Line of Business applications. So we have developed a maturity model for how Cloud Computing will be adoption in different phases. Today the adoption has started with small and departmental applications experimenting with Cloud Services. In the next phase this will move to a Hybrid model, where the Cloud Services will integrate with Data Center applications or services for infrastructure capabilities like Security. This will mature to integrating with other Data Center based business applications. And in the final phase of model, Line of Business applications will migrate to the cloud tipping the adoption of Cloud platforms. From a client perspective as well, the application models will mature from today where they are only Web based applications to a combination of Web and rich client applications and finally reaching the state where Rich, Internet, Mobile or any other future applications types are all first class application models. 7.2 Is Cloud Computing Enterprise Ready? As highlighted in the Cloud Maturity Model, Enterprises will start with small new departmental level applications or web applications that need to scale over time for Cloud adoption. These will also include a class of applications that are already external focused. The key requirements of these applications are abilities of rapid application

353 ISSN Online: development and the elastic nature of the Cloud. There are hurdles that need to be scaled for the successful adoption of Cloud at a level where it can be declared mainstream. At the base level, it is winning the mindshare and driving businesses to even experiment in this area. Issues like lack of Control, Security and Data Privacy, Service Level Agreements, Compliance, Data Loss and new application model costs/adoption are some of the requirements that need to be addressed. While it might seem that these are large and hairy issues, Cloud Providers are working to address these in collaboration with the eco-system. And there are already some shining examples of Cloud adoption within the Enterprise that can help set the tone for the future. 8. IT and Apparel/Garment Industry: The value of information has been well established over the past 20 years. Businesses have long recognized that flawless information flow and knowledge processing streamline business objectives and execution plans, thereby enhancing overall business performance. As businesses entered the era of Information Technology (IT), they began to utilize advanced technologies, such as ERP and CRM systems, to automate information flow and business processes within the company. While up until 5-10 years ago the focus of information dissemination was internal, such a strategy was not a panacea in the growing global marketplace; globalization has forced businesses to work closely with suppliers and other business partners many of them located in other countries or continents to serve customers around the world. Lee et al. (1997) observed that information across different stages of supply chains tends to be distorted, and such distortion leads to poor inventory and production decisions, a phenomenon known as the bullwhip effect.3 Naturally, the need for more data transactions and frequent technology upgrades evolved, and companies started using Business-to-Business (B2B) solutions to automate information exchange between trading partners and collaborators in their business networks. Adoption of such B2B solutions had a positive impact on companies performance, as is evident from a study which was published in early Although some companies were successful in implementing the B2B infrastructure themselves, others found in-house implementation too difficult to manage and principally burdensome for those trying to concentrate on their core competencies. To overcome this problem, some companies started handing over portions of their noncore IT capabilities to external service. Advancement in IT and acceleration of globalization has created another problem for business communities: supporting interoperation of various data formats and communication protocols used by different trading partners. Business communities understand the importance of standard communication protocols and unified data format which not only automate and speed up information transactions, but also enrich the quality of information flow. 9. Functional Framework for Cloud Computing Adoption. 9.1 Business Priorities First of all, companies in India should prepare the business objectives and priorities for adopting Cloud Computing for supply chain management. Majority of the companies are focusing on the priorities most effective and useful aspects of supply chain parameters for attaining maximum benefit from the implementation of Cloud Computing for SCM. The most common priorities are supply chain collaboration to attain collaboration among the supply chain partners. Information sharing capabilities and prospective information reliability is the next priority for some companies where collaboration information is vast and confidential. In addition to this Indian companies are aiming at cost reduction and saving in all aspects of operation of supply chains. And in the light of globalization forces of emerging economy, Indian companies are focusing at achieving greater customer satisfaction through faster deliveries.

354 ISSN Online: Fig3. Benefits of Cloud Computing. 9.2 Cloud Computing Adaption Frame work: The textiles and apparel sector has long been characterized by an elevated degree of complexity, which is an inevitable part of the framework in which firms have to operate. This complexity can partly be traced to behavioral patterns influencing the final purchaser s buying and consumption models, which cause considerable difficulty in forecasting demand when defining apparel collections, and partly also to the short life cycle of a typical garment. The entire sector is affected by these complex interactions, which have repercussions on the strategies adopted by firms seeking to defend a competitive position, as it can be difficult to create and maintain a sustainable competitive edge in an environment where, among other things, an elevated number of production alternatives can be found. Business priorities Online Collaboration Information Sharing Lower Costs 24X7 Support Reliability, Scalability and Sustainability Pay as You Use Lower Capital expenditure Manage Design Implement Benefits Increased Customer Value Agile Deployment Secure Storage an d Management Faster Deliveries Evaluate Fig(3) Cloud Computing adoption frame work. This framework is further complicated by the process of modernization of the distribution network that has taken place in recent years. With regard to industrial organization, in fact, in the Indian context a striking feature is the considerable number of textile and apparel firms, most of which are fairly small, often bound by local aggregations corresponding to the model of the industrial district. Also, as regards the structure of distribution, independent and traditional retailing maintains the largest market share, in

355 ISSN Online: contrast to the typical model in the other European countries, where one finds an increasing predominance of large distribution chains and specialized chains. The frame work suggested here is to be implemented with a strategic approach through the following modules viz., Selection, Manage, Design, Implement and Evaluate. important character for adaptability in emerging economic perspective. The principles, practices, and methods required to raise employee awareness about basic information security and train individuals with information security roles to increase their knowledge, skills, and abilities Selection Selection is based on risk minimization in the following aspects to avoid risk in the following phases or steps of Cloud Computing enablement/adoption. Selection of appropriate Cloud Computing infrastructure and architecture, Selection of right tools matching organizational business priorities Implementation strategies and organization transformation, Meeting current and future needs of Cloud Computing, Immediate updates and technological renovations, Investment decisions and returns on investments (ROI), Disruptions and visibility constraints and obstacles, Documentation and ownership, Education and training to the personnel including the partners, and Overall after performance of the firm. Businesses have to prepare efficient internal systems using Cloud Computing infrastructure to respond quickly to customer s requests, questions, and comments. The operational excellence model for Cloud Computing assisted SCM is that which delivers the highest customer satisfaction on an e-business infrastructure for an emerging economic situation that has the following characteristics; User-friendly, Functional, Reliable, Cost effectiveness, adequacy and Performance. The cost effectiveness is very Manage Identify business priorities and requirements and establish enterprisewide policy for the IT adoption strategy for management of Supply Chain. Acquire and manage necessary resources, including financial resources, to support the Cloud Computing adoption Set operational performance measures for impact of Cloud Computing modules in business operations and metrics like profitability, ROI etc., Ensure the organization complies with Cloud Computing enablement environment Ensure that appropriate changes and improvement actions are implemented as required to adopt Cloud Computing Design Develop the implementation strategies and policies for Cloud Computing enablement Develop administration change management procedures to ensure Cloud environment policies and controls remain effective following a change Define the goals and objectives of the Cloud Computing in the Supply Chain operations like collaboration, Information Sharing etc., Establish a tracking and reporting strategy for Cloud Computing enablement Establish a change management process to ensure transformation of business environment Develop a collaboration strategies with the supply chain partners

356 ISSN Online: Implement Perform a needs assessment to determine risks and identify critical needs based on mission requirements Develop new or identify existing improvement opportunities that are appropriate and timely Communicate management s commitment, and the importance of the Cloud Computing enablement and implementation to the workforce. Ensure that Cloud Computing systems operations and maintenance enables day-to-day business functions Evaluate Collaborate with technical support, incident management, and engineering teams to develop, implement, control, and manage new Cloud Computing enabled SCM administration technologies Assess and evaluate the Cloud Computing security awareness and training program for compliance with corporate policies, regulations, and laws (statutes), and measure program and employee performance against objectives Review Cloud Computing security awareness and training program materials and recommend improvements Assess the awareness and training program to ensure that it meets not only the organization s stakeholder needs, but that it is effective and covers current Cloud Computing security issues and legal requirements Ensure that information security personnel are receiving the appropriate level and type of training Collect, analyze, and report performance measures. 10 Benefits Benefits from Cloud Computing frame work implementation High Level computing Improved Information Sharing capabilities Enhanced Operating Effectiveness Increased Customer Responsiveness Decreased Supply chain complexity 24X7 Support, Pay as you Use Improved Visibility and coordination Improved ROI Reduced stock-outs Optimized inventory Improved Sales Table2. Benefits of Cloud Computing frame work implementation

357 ISSN Online: Conclusions: Depending on the business need, an organization can choose to move certain aspects to their IT requirements to the Cloud Computing. With the correct assessment of the business needs, existing infrastructure and through understanding of an organization s strategic objectives, a capable partner can provide relevant and focused solutions. However, there are a few major factors that hold back business from deploying Cloud Computing. The most prominent one is the security issue. Many potential users are still apprehensive about releasing their in-house data to the datacenter of an external Cloud Services Providers due to issues such as privacy, security etc. Apparel and Garment Companies should develop an overall understanding of Cloud Computing enabled supply chain. Supply Chain Infrastructure and architecture to create Supply chain vision is to be arrived at before implementing the Cloud Computing modules for management of supply chains. In addition, companies should constantly re-evaluate and improve management of supply chains by creating benchmarking efforts in Cloud Computing enabled SCM. [6] Gereffi, G., 2001, "Global sourcing in the US apparel industry", Journal of Textile and Apparel, Technology and Management, 2, 1: 1-5. [7] Amazon Elastic Compute Cloud (EC2): node= , accessed Dec [8] Christopher, M., (2000), The Agile Supply Chain : Competing in Volatile Markets, Industrial Marketing Management, Vol 29, pp [9] Christopher, M. and Towill, D., (2001), An Integrated Model for the Design of Agile Supply Chains, International Journal of Physical Distribution & Logistics Management, Vol. 13, No. 4, pp [10] Hunter NA (1990), Quick Response for Apparel Manufacturing, Textile Institute, UK. [11] Harrison, A., Christopher, M. and van Hoek, R. (1999), Creating the Agile Supply Chain, Institute of Logistics & Transport, UK [12] Johnson, E., (2002), Product Design Collaboration : Capturing Cost Supply Chain Value in the Apparel Industry in Achieving Supply Chain Excellence Through Technology, Vol. 4, Montgomery Research Inc., San Francisco, USA References: [1] ng [2] p [3] [4] Cloud Computing Implementation, Management, and Security John W. Rittinghouse, James F. Ransome, CRC Press. [5] Chan, F.T.S. (2003). Performance measurement in a supply chain. The International Journal of Advanced Manufacturing Technology, 21, [13] Lowson RH, King R and Hunter NA (1999), Quick Response: managing the supply chain to meet consumer demand, John Wiley & Sons: Chichester. [14] Lee, Y. & Kincade, D. (2003). US apparel manufacturers company characteristic differences based on SCM activities. Journal of Fashion Marketing and Management, 7(1), [15] Mustajoki, J. and Hamalainen, R. (2000). Web-Hipre: Global Decision Support by Value Tree and AHP Analysis. INFOR, 38(3), [16] Onesime, O.C.T, Xiaofei, X. & Dechen, Z. (2004). A decision support system for supplier selection process. International

358 ISSN Online: Journal of Information Technology and Decision Making, 3(3), [17] Romano, P. & Vinelli, A. (2004). Quality management in a supply chain perspective. International Journal of Operations and Production Management, 21(4), [18] Chopra, S., and Meindl, P. (2001) Supply Chain Management: Strategy, Planning, and Operation, Prentice-Hall, Inc, Upper Saddle River, NJ. [19] Mentzer, J. T., Foggin, J. H., and Golicic, S. L. (2000) Collaboration: The Enablers, Impediments, and Benefits, Supply Chain Management Review. [20] Mata, F., Fuerst, W. & Barney, J. Information technology and sustained competitive advantage: A resource-based analysis. MIS Quarterly, 19, (1995) [21] Clark, T. & Stoddard, D. Interorganizational business process redesign: Merging technological and process innovation. Journal of Management Information Systems, 13, 2, (1996) [22] Mason T. (1996), Getting Your Suppliers on the Team, Logistics Focus (4:1), pp [23] Beamon, B.M. (1999). Measuring supply chain performance, International Journal of Operations and Production Management, 19 (3), pp [24] Lambert, D.M. and Pohlen T.L. (2001). Supply chain metrics The International Journal of Logistics Management, Vol. 12 No. 1, pp

359 ISSN (Online): Retrieval of average sum of plans and degree coefficient between genes in distributed Query Processing Sambit Kumar Mishra 1, Dr.Srikanta Pattnaik 2 1 Associate Professor, Department of Computer Sc.&Engg. Ajay Binay Institute of Technology, Cuttack, Orissa, India 2 Ex Professor, U.C.E., Burla Director, InterScience Institute of Management & Technology, Bhubaneswar, Orissa, India Abstract Distributed query is one that selects data from databases located at multiple sites in a network and distributed processing performs computations on multiple CPUs to achieve a single result. Query processing is much more difficult in distributed environment than in centralized environment because a large number of parameters affect the performance of distributed queries. The goal of distributed query processing is to execute queries to minimize the response time and to minimize the total communication costs associated with a query. In addition, when redundant data is maintained, one also achieves increased data reliability and improved response time. In this paper, the multi attribute based mechanism is proposed to meet the demand and the result is compared with some commonly used query optimization algorithms. Key words : Query optimization, Plan, Genetic algorithm, Gene, multicast optimization. 1. Introduction The use of a relational query allows the user to specify a description of the data that is required without having to know where the data is physically located. In a relational database all information can be found in a series of tables. The most common queries are select projectjoin queries.minimizing the quantity of data transferred is a desirable optimization criterion. The distributed query optimization has several problems related to the cost model, larger set of queries, optimization cost, and optimization interval. The query optimization process can be defined as follows.

360 ISSN (Online): (i) (ii) (iii) (iv) The input query is fed to the search space and transformation rules are applied to it. Equivalent query execution plan is generated and passed to the search strategy. Formulation of various cost models are done. Best query execution plan is obtained. Considering new large scale database applications, it is necessary to be able to deal with larger size queries. The search complexity constantly increases and makes higher demand for better algorithm than traditional relational database queries. There are number of query execution plans for distributed database such as row blocking, multicast optimization, multithread execution, joins with horizontal partitioning and Semi joins. An optimizer cost model includes cost functions to predict the cost operators, and formulae to evaluate the size of results. Cost functions can be expressed with respect to either the total time, or the response time. The total time is the sum of all times and the response time is the elapsed time from the initiation to the completion of the query. In parallel transferring, response time is minimized by increasing the degree of parallel execution. This does not imply that the total time is also minimized. On contrary, it can increase the total time by having more parallel local processing and transmissions. Minimizing the total time implies that the utilization of the resources improves, thus increasing the system throughput. The main factor affecting the performance is the size of the intermediate relations that are produced during execution. When a subsequent operation is located at a different site, the intermediate relation must be transmitted over the network. It is of prime interest to estimate the size of data transfers. The estimation is based on statistical information about the base relations and formulae to predict the cardinality of the results of the relational operations. The main factor affecting the performance is the size of the intermediate relations that are produced during the execution. When a subsequent operation is located at a different site, the intermediate relation must be transmitted over the network. It is of prime interest to estimate the size of the intermediate results in order to minimize the size of data transfers. The estimation is based on statistical information about the base relations and formulae to predict the cardinality of the results of the relational operations. In this work, we are concerned with the average sum of the plans and degree coefficient between the tasks within the plans in distributed query processing. The structure of the paper is as follows. Review of Literature has been mentioned in Section 2, Problem analysis has been described in Section 3, Need and necessity of genetic algorithm is discussed in Section 4, Problem formulation has been discussed in section 5, Experimental results, analysis and algorithm have been discussed in Section 6, Tables and figures are mentioned in section 7, Discussion & Conclusion has been given in Section 8. and References have been furnished in Section Review of Literature S.Babu et.al [8] have discussed in their paper that distributed database systems use a query

361 ISSN (Online): optimizer to identify the most efficient strategy called plan to execute declarative queries. For a query on a given database and system configuration, the optimizer s plan choice is primarily a function of the selectivities of the base relations participating in the query. Query optimizers often make poor decisions because their compile time cost models use inaccurate estimates of various parameters. Stefan Berchtold et.al [9] have described in their paper that the problem of retrieving all objects satisfying a query which involves multiple attributes is a standard query processing problem prevalent in any database system. The problem especially occurs in the context of feature based retrieval in multi databases. Falout C. Barber et.al [1] have mentioned in their paper that the cost function in task allocation is sum of inter processor communication and processing cost that are actually different in measurement unit. Hong Chen et.al [3] have discussed in their paper that the multi query processing takes several queries as input, optimizes them as a whole and generates a multi query execution strategy. Cristina Lopez et.al [5] have defined in their paper that population of individuals known as chromosomes, represent the possible solutions to the problem. These are randomly generated, although if there is some knowledge available concerning the said problem, it can be used to create part of the initial set of potential solutions. 3. Problem Analysis The query is submitted by user to the query distributor agent and then it will be distributed.after receiving the user query, the query distributor agent sends sub queries to responsible local optimizer agents. The query distributor agent can also create search agents if needed. The local optimizer agents apply a genetic algorithm based sub query optimization and return a result table size to the global optimizer agent. The global optimizer agent has the responsibility to find best join order via network. The global optimizer agent receives resultant table size information from local optimizer agents. Using an evolutionary method, it finds a semi optimal join order. However, this time the genetic algorithm fitness function is based on minimizing communication rate among different sites. 4. Need / necessity of Genetic algorithm In distributed query processing environment, a single query may have a single plan or multi plans. Similarly a single plan may have a single task or multi tasks. In the same manner the multi query may have multi plans with multi tasks per plan. In this work our aim is to find average sum of the plans and degree coefficient between the tasks within the plans. Since it is a NP complete problem we need Genetic algorithm to solve the problem. 5. Problem Formulation The first step to represent this problem as a genetic algorithm problem is determining the chromosome, genetic algorithm operators and fitness function. For the crossover, one point in the selected chromosome would be selected along with a corresponding point in another chromosome and then the tails would be exchanged. Mutation processes causes some bits to invert and produces some new information. The only problem of mutation is that it may cause some useful information to be

362 ISSN (Online): corrupted. Therefore the best individual is used to proceed forward to the next generation without undergoing any change to keep the best information. Probability for mutation operation=0.001 Crossover point= round(1+rand*(size of chromosome 1)) Defining fitness function is one of the most important steps in designing a genetic algorithm based method, which can guide the search toward the best solution. After calculating the fitness function value for each parent chromosome, the algorithm will generate n number of children. The lower a parent chromosome s fitness function value, the higher probability it has to contribute one or more offsprings to the next generation. After performing operations, some chromosomes might not satisfy the fitness and as a result the algorithm discards this process and gets q (q<=n) children chromosomes. The algorithm then selects n chromosomes with the lower fitness value from the q+n chromosomes ( q children and n parents) to be parent of the next generations. This process will be repeated until a certain number of generations are processed, after which the best chromosome is chosen. Sum=0; 6.1. Algorithm to=initial CPUtime; t1=cputime after mutation and crossover operation CPUtime, t2=t1 t0; for i=1 : number of queries planselect(i)=x(i)/(numberofqueries*planquery) real_cost(i)=planselect(i)/numberofqueries +t2; est_cost(i)=real_cost(i)/number of queries; sum=sum+est_cost(i); 6. Experimental results, analysis and algorithm Maximum generations=100 Numberofqueries=100 Number of relations=100 Size of Chromosome ( Plan in a query)=5 Population=round(rand(numberofqueries, sizeof chromosome)) Probability for crossover operation=0.06 end avgsum=sum/number of queries; for i=1 : number of queries if(avgsum>est_cost(i)) geneval(i)=(avgsum real_cost(i))+est_cost(i); else geneval(i)=(real_cost(i) avgsum)+est_cost(i); end end

363 ISSN (Online): x(i) represents number of chromosomes. As shown above in the table, we have 10 different plans of various sizes. Crossover point, cp=2 Size of chromosomes=5 There is no doubt that dynamic programming methods always give us optimal solution. However, since the time and space complexity of the genetic algorithm base optimization is much less, it is not a practical approach for high amount of nested joins. An evolutionary query optimization mechanism in distributed heterogeneous systems has been proposed using genetic algorithm approach. Genetic and randomized algorithms do not generally produce an optimal access plan. 7. Tables, Figures For example, the estimated cost of 5 th plan is The value of the plan is 11. The corresponding cost of Gene value in this case is The estimated cost of 10 th plan is The value of the plan is 19. The corresponding cost of Gene value in this case is The CPU time recorded after mutation and crossover operation is The average sum of plans is The average association degree coefficient between genes is Table 1 Plan Populat ion X(Plan) Est_cost Gene value Figure Plan ( Plan Vs Cost of Gene value) 8. Discussion and Conclusion

364 ISSN (Online): Data allocation defines the type of data stored while operation allocation states where accessing and processing of operations i.e. Select, Project, Join etc. are taken place. The problem of retrieving all objects satisfying a query which involves multiple attributes is a standard query processing problem prevalent in any database system. Task allocation is an essential phase in distributed database system. To find the solution to task allocation problem, complete knowledge about the tasks and processors should be accumulated. 9. Reference [6] W.T.Balke and V.Guntzer, Multi objective query processing for database, VLDB, [7] K.Gajos and D.S.Weld, Preference elicitation for interface optimization, UIST,2005. [8] S. Babu, P.Bizarro, D.Dewitt, Proactive Reoptimization Proc. Of ACM SIGMOD Intl.conf. of Management of Data, June 2005 [9] Stefan Berchtold, Cristian Bohm, University of Munich, Oettinge Str, Germany [1] Falout C. Barber, R.Flickner M, Efficient and effective query, Journal of Intelligent Information Systems [2] Guohua, Shuzhi Zhang, Dongming Zhang, The College Of Information Science and Engineering, Yanshan University, QinHuangdao, China International Conference in Computational Intelligence for modeling, IEEE transaction, [3] Hong Chen, Sheng Zhou, Shan Wang, School of Information, Remin University China, International Conference in Data and knowledge Engineering, [4] Lynda Tamine, Claude Chrisment, Mohand Boughanem, University Of Toulouse, France, IPM [5] CristinaLopez, Vicente P.Guemero Bote,University of Extremadura,Badajoz, Spain Proc.of ACM Sigmod Intl.conf of Management of Data, June First Author: Sambit Kumar Mishra 1. Passed B.E. in Computer Sc.&Engg. From Amravati University, Maharastra in Passed M.Tech. in Comp.Sc. from Indian School of Mines, Dhanbad. 3. Continuing ph.d.(comp.sc.)in Optimal Query Processing Techniques using soft computing Tools under Prof.Dr. Srikanta Pattnaik, Ex Professor, U.C.E., Burla. 4. Total Teaching Experience : 16 Years in various Engineering Colleges, Orissa, India. 5. Life Member of Indian society for Technical Education. 6. Member of International Association of Engineers. 7. Participated and Published 07 nos. of Conference papers in National and International Conferences. Second Author: Dr.Srikanta Pattnaik 1. Passed B.E. from U.C.E., Burla. 2. Completed M.E. and ph.d. from Jadavpur University, West Bengal. 3. Guided more than 07 ph.d. students. 4. Editor and Editor in chief of many journals.

365 ISSN (Online): A Goal Based Approach for QFD Refinement in Systematizing and Identifying Business Process Requirements Atsa Etoundi Roger 1, Fouda Ndjodo Marcel 2, Atouba Christian Lopez 3 1 Department of Computer Sciences, University of Yaoundé I, Yaoundé, Cameroon 2 Department of Computer Sciences, University of Yaoundé I, Yaoundé, Cameroon 3 Department of Computer Sciences, University of Yaoundé I, Yaoundé, Cameroon ABSTRACT The traditional Quality Function Deployment (QFD) methodology has successfully been used in many organizations in order to increase the productivity and the quality of service. However, this methodology has many limitations such as the ambiguous and unsystematic identification of customer requirements. In this paper, based on a goal oriented approach for the definition of business process requirement model, we define a model that overcomes the defined limits. Keywords: Business Process Modeling; QFD, Requirement Engineering; Software Component; Requirement Relevancy; Application Engineering; Requirement representation 1. INTRODUCTION The management of the quality of service in different enterprises is problematic and is a challenge in the network economy. In this area, only organizations which design their business processes based on the satisfaction of customer s requirements will survive. For this purpose, many organizations are looking for methodologies that can help in identifying the needs of their customers. In the new economy, the gap between producers and consumers is required to be minimized since they have work hand in glove to deal with the pressure of the network economy. As highly, customized products and services replace mass production, producers must create specific products that are imbued with the knowledge, requirement, and tastes of individual customers. The consumer has become involved in the design process of goods and services for his satisfaction. The main objective for enterprise managers is to identify requirement amount respective customers for the delivery of qualified outcomes. The difficulties in doing this are to get the maximum attributes from various clients and to define the degree of importance of these attributes. In many enterprises, the Quality Function Deployment (QFD) is used as a tool for improving the development and manufacturing products that better match customer expectations [1][6][7].The QFD is a methodology that is based on four phases and uses a matrix to translate customer requirements from initial stages through production control. Amount the phases is (a) the documentation of customer requirements, (b) the creation and specification of product concepts, (c) the definition of the process planning and documentation of process target values, (d) the creation of the product performance indicators for the monitoring of the production process. Despite the popularity of this tool, managers still have difficulties in systematically identifying customer requirements and defining the degree of importance associated to each requirement. This paper presents a Goal oriented Approach for the definition of a business process Requirement Model (GAReM) to overcome the above limits. The rest of the paper is structured as follows, in section 2, we defined basic concepts that form our Goal oriented Approach to identify and represent customer requirements. In section 3, we present our model for requirement identification and importance. In section 4, we present the proposed model, and in section 5, we conclude the paper and highlight some future works. 2. BASIC CONCEPTS Our concern, in this section, is to define the concepts necessary in comprehension of the formal representation of a need expressed by a user.

366 ISSN (Online): User Requirement 2.3 Domain We call requirement of a staff in an organization, an activity in the business process of the said organization, assigned to this staff and of which he would like to have automated in the future computing system. To model a need, it should be identified as a prerequisite. This is why we propose the elaboration of a form to identify user requirements. 2.2 Eliciting of Requirements The document for requirement eliciting is a form to be submitted to users in order to collect from them, the literal description of their expectations. In each form, only an expectation of a staff is described. The form in question is structured as follows: Structure or service : describes the structure of the organization, where the staff is assigned; staff : represent the name of the staff who is expressing the requirement; priority of expectation : the priority of expectation associates a level of importance to that expectation; Goal : the usage intention of the user; rules : represent the business rule of the intention of use; Constraints: constraints indicate nonfunctional expectations which could impede the realization of user goal. They could be linked with the man-machine interface, security, etc; Domain of expectation: domain of expectation describes the context in which the user usage intention is expressed. This field is filled by the software engineer; Status of expectation: it indicates the expectation state. it can take one of the following values: proposed, rejected, validated, taken into account. The inventory of user needs makes it possible to identify in an exhaustive way user expectations. Expectations listed at the level of the users are by default in a proposed state. They pass to a rejected state if user expectations were not approved by a senior staff in rank and if not validated. The taken into account state indicates that an expectation was taken into account in the user requirements specification model. In a formal way, we will represent a user requirement by where denotes the structure to which the user is assigned, denotes the name of the user who expressed an expectation, denotes the level of importance of the user usage intention the users usage intention, business constraints, the domain of the expectation, and the status of an expectation. A domain is the field of application of the users usage intention or simply the context in which an expectation is defined or applies. 2.4 Constraints The constraints indicate the non-functional expectations which could impede the realization of the users goal. It represents the state of the environment in which the task will be carried out. 2.5 Goal Definition 1: functional goal A functional goal defines a requirement, expectation that the system can satisfy, it expresses what the user of the system would wish to do [10]. We deduce from this definition, that a functional goal is simply a text expressing the usage intention of an unspecified user. This implies the use of the active voice and action verbs. Consequently, there is existence, in the text, of a group of words translating the action which is likely to be accomplished by the user on the one hand, and on the other hand, the existence of a group of objects which are subjected to the action of this user; and if possible, the anticipated end result. It is stated in [15] that an action verb expresses what the subject does or is subjected to and [16] informs us that the infinitive is the simplest form of a verbal expression. We shall chose to represent an action translated by the group words previously mentioned, by an action verb in its infinitive form. This gives us the following representation of the goal: Rule (1) : Goal: in the + Article + Name of (1) by: Action verb infinitive Possible articles are: either the definite articles (the) or indefinite article (a, an); Name of object or nominal group which represents either a set of entrants of the same nature as the system; either a set of artifacts of the same nature as the system. It is derived from the name of knowledge bits which encapsulates the goal. In a formal manner, a goal b will be represented where: Object

367 ISSN (Online): Definition 2 : consistent goal We will say that a goal is consistent when its expression is in conformity with the rule (1). Definition 3 : inconsistent goal A goal is inconsistent when its expression is not in conformity to the rule (1). Lemma (1): Any inconsistent goal can be transformed into a consistent goal. Proof: By definition, a goal is a text translating the usage intention of a user. From [13] and [12] this text is compressible into a text representing the intention in its entirety. As we are in presence of a text translating an intention of use, so we can identify the action and the object on which this action applies. Given that the action is known according to what precedes then, it is trivial to find the verb that translates this action, on the one hand and on the other hand find the article of the group of objects. The transformation process of an inconsistent goal into a consistent goal is called the globalization of the goal. Let b be an inconsistent goal, and a function :, the globalization function. represents the consistent goal deduced from the globalization of b. is called globalized goal or global goal of b. Applying [12] and [13], globalizes the usage intention of the user. This means that, it presents the users usage intention in its entirety. Thus, we can deduce that the intension of the user is a refinement of the globalized goal. We deduct from this fact that a globalized goal always admits at least one sub goal that represents the user's specific need. A goal that does not admit any sub goal is called: an elementary goal. Let's denote a set of consistent goals of an organization by ;, the set of intentions expressed by the staff of this organization; and f a function defined as follows:, In this approach the goal is the key concept. In the rest of this section we shall give definitions of the other fields of the knowledge bits. These other fields will have their importance when it will be necessary to pass our specification to other specifications. 2.6 Business Rules Definition 5: business rules Business rule defines a law in the domain to which the goal must conform itself. It helps in the setting up of a management process to achieve the goal. Business rules have an impact on the external environment (the organization), as well as on the internal environment (the system) [17]. They are tree types: scheduling rule, trigger rule or static constraints [10]. A scheduling rule is a law of the domain that describes the order in which goals must be achieved, for an instance of a given object. The trigger rules and the static constraints keep the same semantics as in [10]. Definition 6 : Polysemous goals A goal is termed polysemous if the business rule that governs it changes from one domain to another. 2.7 Importance of an Expressed Requirement The level of importance of the expressed requirement expresses the credit that a user gives to this expressed requirement. It can take one of the following values: very important and indispensable, important and indispensable, indispensable, important, necessary, and in bonus. It is possible to choose a numeric value to express the level of importance of a user's expectation. We noted through a survey carried out on a sample of twenty-five Cameroonian government services that the staff had difficulty assigning some numeric values to materialize the credit associated to their expectations. To this effect, we recommend the use of an assessment scale level of importance of requirements according to the target population (futures users of the system) 2.8 Knowledge Bits or Expressed Requirements An expressed requirement or knowledge bit is defined in the following manner:,,,, where represents the name of a concept of the domain. In the definition [10], we replaced the conceptual pieces by constraints, level of importance of the user usage intention, because in the course of this work we shall be interested exclusively in the formalization of the representation of user requirements. In the intention to integrate the level of importance of user goals, we introduced the field level of importance in the knowledge bits. The constraint field was added to exhibit the non functional requirements which are hidden behind the usage intention of the user. These non functional needs could be refined by software engineers. Constraint: - The eligible action verbs henceforth shall be those that translate actions, by preference management actions, susceptible of being automated by a

368 ISSN (Online): computer system. For example, verbs: to speak, to eat, to jump, to run, to laugh, to sell, to paint, etc.., are excluded; - our work is exclusively about abstract goals; - goals expressed by sentences in the active voice. We have defined the basic concepts necessary to understand our approach. In the following section we are going to formalize the above mentioned concepts. 3. GOAL ORIENTED APPROACH FOR THE DRAWING UP OF REQUIREMENT MODEL 3.1 Eliciting User Requirements The Elicitation of user requirements is an activity aimed at collecting user needs, as well as validating user requirements. With this intention, the software engineer must use a form structured as the requirement eliciting document as defined in section 2 paragraphs 2.2. The software engineer should first have the different identified requirements validated by a competent staff of the organization in the presence of the users who initially expressed the requirements. Once validated, the software engineer must complete the domain field in each form. As we proceed, only validated requirements shall henceforth be considered. We shall then indicate user expectations or expressed requirements; requirements whose state is validated, Ω will represent the set of expressed requirements of the organization. Let's consider Ω,. will represent the field of 3.2 Selection of Requirements Let's consider a human language (French, English, German, etc. ), we define the function as follows:,,,,, The symbol shall be used to express negation, represents an expression opposite in meaning to. Let's consider a and b two requirements of the organization, with a and b elements of Ω. We have:,,,,,,, and,,,,,,,. - Property 1 (inconsistent requirements): will be said inconsistent if and only if. is inconsistent. - Property 2 (ambiguity of requirements): a and b will be said ambiguous if and only if :.,.. and... - Property 3 (similarity of requirements): and will be said similar if and only if:.. and... - Property 4 (contradiction between requirements): and will be said to be contradictory if and only if at least one of the following conditions are satisfied.... ; Property 5 (identity of requirements): and will be said to be identical if and only if : Property 6 (Consistency of requirements): is said to be consistent if and only if for any requirement b none of the above properties is satisfied. Let's consider as set of retained requirements, and R those rejected. The process of requirement selection consists in: 1-, 2- Ω ;, 3- if verifies the property 1 : rejected a ; Ω Ω ; 4- if there exist and verifying the property 2 to 4 : reject and ; Ω Ω, ;, 5- if there exist and verifying the property 5 : Ω Ω, ; 6- if verifies the Property 6 : Ω Ω ; Repeat steps 3 and 6 until Ω. Elements of R must be the subject of discussion with the staff who expressed them. At the end of discussions repeat steps 2-6. This activity aims at discovering and deleting any requirement with an unnecessary user goal. The software engineer must rely on his understanding of the different usage intentions of the user. When one of properties 2-7 is verified, it is recommended to discuss with those who expressed these intentions in question. It is only if the latter is confirmed to be of same intention that this group is validated. At the end, the software engineer summarizes the requirements by goal and submits them once more for validation by the organization. This grouping ensures that the semantics of goals is the same for everybody (users and software engineers) and that there is no double use of a goal. This is why the classification of user requirements is done by goal. It enables us to detect cases of double use. 3.3 Transformation of Requirements into Knowledge Bits Let s consider a requirement a of Ω,,,,,,,,, and a Knowledge bit, by definition, C can be expressed in the form

369 ISSN (Online): ,,,,. Our objective is to be able to construct from. From rule (1), we have :.,,. Rule 2 (translation of requirements):.. ;. ;. ;. ;. ; and.. This activity aims at transforming users requirements into knowledge bits, integrating the level of importance of each expressed requirements. This procedure is repeated for each element in Ω. Knowledge bits from the translation form the set of knowledge bits of the organization. 3.4 Development of Requirement Model Problem Frames A problem frame [18] (or problem diagram) is a diagram that defines in an intuitive manner a class of identified problems in terms of its context and domains characteristics, interfaces and requirements. The system to be developed is represented by a machine. For each problem frame, a diagram is established. Simple rectangles denote application domains (which already exist), rectangles with a double bar denote domains machine which are to be realized, and requirements are noted by an oval dotted line. Lines joining them represent interfaces, also called shared phenomena. Jackson distinguishes causal domains which obey certain laws, lexical domains which are the physical representations of data, and domains biddable (give out orders) which are people. The use of a problem diagram consists in instantiating domains, interfaces and requirements. In the continuation we shall rely on problem diagram concepts to do demonstrations. We will designate by machine, the machine presented earlier. This machine transforms knowledge bits (problem of the real world) into future exigencies of the system; shall represent the space of knowledge bits (problems) of the organization and problem frame of the organization, it is the set of exigencies of the organizations future system. S will represent the set of all objects of the organizations information system. The latter are expressed in business rules Basic Axioms for the Development of the Requirement Model of an Organization Let s consider a knowledge bit. element of,,,,,, a machine as expressed in [18] and an object of. We note :,.,.,.,., The processing of the object by the machine W, in the context., under the rule. and the constraints., such that the goal. is satisfied. We shall say such a machine recognizes requirement. We construct in the following manner:,,.,.,.,.. is the set of objects of the organizations computer system for which the expectation. is satisfied under the rule. and the constraint.. next, the notation,,,, shall be replaced by,,,, and by ; the knowledge bit shall be called requirement or expectation ; shall represent the set of exigencies of the system (software requirement) which are satisfied in the context., under the rule. and the constraint. ; is use to represent an undetermined value of a field Axiom 1 : Coherence of Knowledge Bits Let s consider an expectation b of, we say that b,ω,,, is a coherent requirement if and only if : (1) -S ; (2) - PF PF Axiom 2 : Concept of Sub-requirement Let and two coherent requirements of and two machines V and W such that V recognizes a and W recognizes. (3) we say that,,,, is a subrequirement of,,,, or that ω is a refinement of ω, if and only if: and, such that W e, b and the execution of Ve, a does not satisfy. (4) We say that b is a generalization of if and only if is a sub requirement of. We note ρ that generalization of. is called the specific goal of ω (5) We say that and of are traceable if and only if : or. (6) A requirement is incomplete if it has only one daughter requirement. Incomplete requirements must be resolved by addition of daughter requirements. If a daughter requirement has a daughter then add other daughter requirements; else merge father and daughter requirements as a unique requirement. (7) if, then is scheduling rule of daughter requirements of Axiom 3 : Merging Requirements Let a et be two coherent requirements of, and,, three machines such that M recognizes and recognizes, (8) We say that,,,, can be merged with,,,, if and only if:

370 ISSN (Online): Ω ω, c,ω,,,, W, b / S S S, e S, d S W e, c and W d, c ; where :, ω is a goal including ω and ω. scheduling rule;, is the highest level of importance between and Axiom 4: Importance of Requirements. Consider two coherent requirements,,,, and,,,, of with and of values taken from an ordered set. We shall say that a is more important than b if and only if : Axiom 5: Ambiguous requirements Two coherent requirements,,,, et,,,, of, are said ambiguous if and only if :, Axiom 6: Partitioning of Requirements Let n coherent requirements,,, of, and a coherent requirements,,,, of, S C such that S, then the requirements h is sub divisible into sub requirements,,, where and Axiom 7: Identical Requirements Two coherent requirements,,,, and,,,, of, shall be said to be identical if and only if :, Can we say that the requirements of an organization are classified in hierarchy? Our objective in this section is, firstly, to adopt a formal proof that requirements as we represent them enable the description of all the activity of an organization; secondly, to formally define when we can consider requirements of an organization as entirely described; and thirdly, to give some characteristics of the set of requirements of an organization Formal Proof on the Completeness of Description of a Business Process To show that our organizations requirement representation model enables the description in an exhaustive manner the requirements of a business process of this organization, it is sufficient to show that this representation is another way to describe a business process. For this we are going to rely on the work of R. Atsa and Mr. Fouda in [21, 22] firstly, we shall construct a requirement from the concept of " task " and its underlying notions; secondly, we are going to show that the properties relative to the description of the business process, elaborated in [21] are applicable by the set of requirements of a business process. R. Atsa and Mr. Fouda defined in [21] the concept of business process of an organization and its underlying concepts. In their vision a business process is a set of tasks that must be realized in a context, for the attainment of objectives or precised goals. A task is seen as data of: a realization context; a set of states; a transition function between states; and of an objective to attain or goal. In this respect the achievement of a goal is realized by the observation of values associated to each indicator, linked to the execution of one or a set of tasks. A state is the set of objects of the organization on which the task acts to achieve fixed objectives Let s consider a requirement,,,, of, and a task t. we have : the set of constraints linked to the execution of t the set of values of indicators from which we observes the achievement of the objective associated to : : the transition function between states associated to the context of execution of the task. the set of objects of the organization on which the execution of acts. a) Let's construct a as a function of t and its underlying concepts: From definition 1, ω is expressed in the form of a triplet, τ, τ under the conditions of section 2.5. express ω as a function of t in the following manner: ωtraiter, obsert, obsert From definition 5, write : f ; and from definition of domain (cf. section 2.3), express : contt ; While relying on the definition of constraints (cf. section 2.4), write: constt ; similarly (cf. section 3.4.1), we express S in a similar manner as follows: S etatst. Without deviating from the general rule, we shall consider the entire task in the approach of [21, 22] have the same importance. To this effect we put 1. is expressed as a function of t in the following manner:,,,,1, where :,, b) Let's show that properties related to our specifications are applicable to tasks as defined by R. Atsa et M. Fouda. In [22], several properties have been elaborated on the set of states of the environment. They induce a set of dependency rules between the different

371 ISSN (Online): tasks of the business process. Key among these are the following: consistent state (Useful State), Equivalence between states (Equivalent of state), sub - states (Substate). Our work shall consist in showing that axiom 1, 2 and 7 are applicable to the concepts of tasks as defined in [21]. Let s consider two requirements and, with and of, and two tasks and. We suppose that is associated to and at : 1 st case (consistent requirement):, from what precedes, ; where, is a consistent state. 2 nd case (identical requirement): we deduce that 3 rd case (sub requirement) :, from what precedes consequently, from what precedes,, where. We have just shown that our requirement modeling approach of an organization is another manner to describe a business process. R. Atsa and Mr. Fouda described the business process from the angle of tasks and states of the environment, in this paper we showed that a description of the business process can also be made from the angle of requirements and goals When can we consider that the requirements of an organization have been described entirely? We are going to consider that needs of an organization have been described entirely when: all traceable requirements are complete, and the associated business rules to these requirements are the scheduling rules. the elementary requirements are constituted of all elementary tasks, there exist no knowledge bits that could either be father of another, nor daughter of another.. business process can be split into tasks and forms a hierarchical set. We showed in section that to each associated task of a professional process is associated a unique requirement. A sub business process is by definition a set of divisible requirements in which represents the tasks of this sub process. 4. The proposed methodology The proposed methodology is presented in figure1. This methodology is based on the notion of forms that are filled by customers of a given enterprise for the satisfaction of their needs. The GAReM defines a methodology that systematizes the identification of customer requirements by eliminating ambiguities that are fund in traditional forms. From the forms that have been filled by customers, the designers team easily identifies and extracts the requirements based on the quality of service needed, and the different values associated to each of the requirements. These values represent the degree of importance perceived by customers. Systematization Customer Requirement Of identification GAReM INPORTANCE Degree of Importance CORELATION MATRIX FUNCTIONS RELATIONSHIP MATRIX TARGET VALUES Figure1: The proposed methodology Customer perception Some characteristics of the set of requirement Lemma (2): The requirements of an organization can be classified on the basis of hierarchy. Proof: Let's consider any business process, according to what precedes; we can obtain the set of requirements which characterize. Let s show that for any element in. Let be either father, or daughter of a requirement. From [19] the 5. CONCLUSION AND FUTURE WORKS Dealing with the quality of service within an enterprise is very challenging in the new economy. Enterprise that fails to take into consideration this new aspect will have a lot of difficulties to survive. Taking into consideration the satisfaction of customers requires a systematic identification and classification of customers needs. The identification and classification of customers requirements is ambiguous and complex as these

372 ISSN (Online): requirements are defined using a traditional approach such as form filling process. In this paper, we have defined a methodology based on the goal oriented approach for the definition of a business process requirement model (GAReM) to systematically identify and evaluate customers requirements. In this approach we give a formal description of a requirement and point out properties that should be satisfied by a given requirement model. For a model of customers requirements in a given organization, we conjecture that these requirements can be classified on the basic of hierarchy in a specific business process. However, the paper does not deal with the representation of business rules and does not highlight the influence of the degree of importance of the requirement in the correlation matrix, and in the customers perception of the quality of service. These limits are some works that can be carried out for the improvement of our modeling approach. 6. REFERENCES [1] Moses L. Singgih, Anggi I. Pamungkas, «Implementing Grey Model and Value Analysis in QFD Process to Increase Custumer Satisfaction (Case Study at Juanda International Airport-Surabaya», 3 rd International Conference on Operations and supply Chain Management, Malaysia [2] Karen Mc Graw, Karan Harbison, User Centered Requirements, The Scenario-Based Engineering Process. Lawrence Erlbaum Associates Publishers, [3] The Standish Group, Chaos. Standish Group Internal Report, [4] Voas J., «COTS Software - The Economical Choice?», IEEE Software, vol. 15 (3), p ,March, [5] Tran V., Liu D.-B., «A Procurement-Centric Model for Engineering Component Based Software Engineering», Proceedings of the 5th IEEE International Symposium on Assessment of Software tools (SAST), Los Alamitos, California, USA, June, [6] Hang-Wai Law, Meng Hua, «Using Quality Fonction Deployment in Singulation Process Analysis», Engineering Letters, February [7]Jacob Chen, Joseph C.Chen, «QFD-based Technical Textbook Evaluation- Procedure and a Case Study», Journal of Industrial Technology, January [8] Güngör-En C., Baraçli H., «A Brief Literature Review of Enterprise Software Evaluation and Selection Methodologices : A Comparison in the Context of decision- Making Methods», Procedings of the 5th International Symposium on Intelligent Manufacturing Systems,p , May, [9] Kahina Hassam, Bart George, Régis Fleurquin, Salah Sadou, «utilisation de la transformation de modèles pour faciliter la sélection de composants logiciels», IDM juin Mulhouse, [10] Farida Semmak, Joël Brunet, «Un métamodèle orienté buts pour spécifier les besoins d un domaine», 23e Congrès INFORSID, pp , mai [11] Bouchra El Asri, Mahmoud Nassar, Bernard Coulette, Abdelaziz Kriouile, «Architecture d assemblage de composants multivues dans VUML», Revue RSTIL Objet, PP 1-32, [12] Minel J.-L., «Le résumé automatique de textes : solutions et perspectives», in Proceedings of TAL, Résumé automatique de textes, vol. 45/1, p. 7-13, [13] Mehdi Yousfi-Monod,Violaine Prince, «Compression de phrases par élagage de leur arbre morpho-syntaxique : Une première application sur les phrases narratives», RSTI- TSI 25/2006. Document numérique, pp , 2006 [14] G. Grosz, «Ingénierie des besoins : problèmes et perspectives», Centre de recherche, Paris Sorbonne 1, [15] François Ott, Pierre Vaast, «livre de grammaire : sixième en troisième», Hatier, 1991 [16] François Ott, Pierre Vaast, «livre de grammaire : Première en Terminale», Hatier, 1991 [17] Zave P., Jackson M., «Four Dark Corners of requirements Engineering», ACM Transactions on Software Engineering and Methodology, 1997 [18] Michael Jackson. «Problem Frames. Analyzing and structuring software development problems». Addison-Wesley, 2001 [19] Jean-Noël Gillot, «La gestion des processus métiers: l'alignement des objectifs stratégiques de l'entreprise et du système d'information par les processus», pp , IDM 1997 [20] Alan W. Brown, Sterling Software et Kurt C. Wallnau, «the Current State of CBSE», IEEE Software, Software Engineering Institue, October 1998 [21] R. Atsa Etoundi, M. Fouda Ndjodo, «Human Resource Constraints driven Virtual Workflow Specification», IEEE SITIS pp , [22] R. Atsa Etoundi, M. Fouda Ndjodo, «Feature- Oriented Business Process and Workflow», IEEE SITIS pp , Roger Atsa Etoundi is a Senior Lecturer in the University of Yaounde I, Cameroon. In 2004, he defended his PhD thesis entitled A Domain Engineering Approach for Multi Perspective Business Process and Workflows Modeling. He has published 12 articles in this field of research. Since 2005 till date, he is currently the Chief Information Officer in the Ministry of the Public Service and Administrative Reform where he engineers the Workflow of the management of different State Personnel. Marcel Fouda Ndjodo is a Senior Lecturer in the University of Yaounde I, Cameroon, and Head of the Computer Science Department in the Higher Teacher s Training Institute (Yaounde,Cameroon). Atouba Christian Lopez is a PhD student

373 351 Application of Some Retrieved Information Method on Internet Vu Thanh Nguyen 1, Nguyen Quoc Thinh 1 1 University Of Information Technology, 34 Truong Dinh Str., 3 Dist., HoChiMinh City, VietNam Abstract This paper compares several methods of information extraction on the internet. Today, internet has become a treasure of knowledge. Every year, thousands of pieces of different information are posted on the internet. So, extracted information on the internet for many different purposes has become an important problem today. Users may extract information based on some available tools such as Lapis, Risk, Rapier, Wien, and Stalker However, these tools have a disadvantage: we must update the training data when the website changes. So SVM and CRF associated with natural language processing are the best solutions to solve this problem. Information extraction from online Vietnamese news website with SVM and CRF brings experiment results that is very optimistic. Its results reach nearly 90% of the accuracy in websites and the processing time is less than one minute per site when the specified number of link levels is 1 within the base site. Keywords: RI (Retrieved Information), CRFs (Condition Random Fields), SVM (Support Vector Machine), ECT (Embedded Catalog Tree). I. Introduction about retrieved information on the internet Nowadays, the internet is a huge library of the whole world, with a lot of information that includes all fields, all jobs in industry, agriculture, economy, finance So this retrieved information has become very imperative. We need to solve the problem: How is realized and retrieved information extracted as exactly as possible from the internet? II. Some methods for solving this problem Hand-code wrapper [5]: the first wrapper was created is hand-code wrapper, this method uses the consistence of web pages to create its wrapper. For example, information from showtimes.hollywood.com can be retrieved easily by a simple structure: information like show-time, movie title always stays in special tags (show-time always is bold with tag <B></B>). The advantage of this method is simple and this wrapper is created easily. However hand-code wrapper is only used to present some consistent web pages and each wrapper is only used for one web page. Nowadays, with the continous changing of web pages, it is useless to use hand-code wrapper, that s also the reason why automatic wrapper construction was taken from. Automatic Wrapper Construction (AWC) [5][6]: With this method, wrapper was created automatically based on some requests of users, such as information on what field, what career, And each wrapper can be used for a group of web pages (with similar structure). AWC-LR Wrapper Class [5][6]: LR (Left Right delimiters) is a set of delimit symbols on left and right side of field of needed information, input for this method is only the address of web pages which have marked the field of needed information, and after that it will try all of cases of delimit to create a suitable wrapper. The advantage of this method is easy to use, we just add addresses of web pages, the rest tasks will be automatic and receive wrapper for web pages. However, this method can only operate perfectly with web pages, where information have consistent delimit symbol and at the same time information is only realized based on LR delimiters, so it is easy to meet an error. For example: <HTML><TITLE>Country Codes</TITLE> <BODY><B>Congo</B> <I>242</I><BR> <B>Spain</B> <I>34</I><BR> <HR><B>END</B></BODY></HTML> We determine (<B>, </B>, <I>, </I>) with LR Wrapper Class. AWC-HLRT Wrapper Class [5][6]: H(Head delimiter), LR(Left Right delimiter), T(Tail delimiters) are delimit symbols at head, left, right and tail of information we want to retrieve. This is an improved model of LR Wrapper Class, so its advantage and disadvantage are the same. There is just a different thing, this method has Head and Tail delimit symbol at bonus, and therefore retrieved information can be marked more exactly. Embedded Catalog Tree (ETC) [5]: Leaves are the items users want, local nodes (even root) represent a list of data sets, each item in any data set may be a leaf l or a list L (Embedded List) of k which lets we know the number of items in data sets. Preparing with the two above methods, LR and HLRT Wrapper Class are more improved with more exact information retrieving. Proceeding from ideas, information in web pages was represented based on level and data sets. When user adds web pages, which field of want-to-retrieve information was marked, with this field and ECT method, a new rule set was born. Retrieving process will execute with path P form root to suitable leaf and retrieve every x which belongs to P from its father node. Rule set was presented by Landmarks, Wildcard, Function, Cascade Function, Selection Rules. Landmarks: group of continuous tags (like works, numbers, HTML Tags, substring, ) for example: <b>,<p>

374 352 Wildcard: class of tags (like number, Sign, HtmlTag, Allcaps, ) Function: an expression with a landmark or a wildcard is variable. For example: SkipTo(:<b>) Cascade Function: like SkipTo(Allcaps), NextLandmark(number) Selection Rules: like SkipTo(<b>) or SkipTo(<i>) state s t-1 and label l of current state s t. Usually, training CRFs is proceeded by maximum likelihood training data along with optimal technology such as L-BFGS. The argument (based on learned model) is finding out correlative label sequence of input observe sequence. With CRFs, usually the use of traditional motion scheming algorithm is Viterbi to argue with new data. Its advantage is that we can retrieve information exactly up to 90%. However, a lot of web pages always change content and way to show information, accordingly wrapper has been updated regularly. This also is mistake of almost wrappers nowadays. III. Retrieved information on news web pages with method combined Condition Random Fields (CRFs) with Natural Language Processing (NLP) CRFs is a model of undirected linear state (a trained limited state engine with conditions) and follows the first Markov property. CRFs was proved successfully for labels assign projects like separate words, assign labels for phrase, determine entities, assign labels for group of nouns, etc Call o = (o 1, o 2,, o T ) a sequence of observed data which will be assigned labels. Call S a state set, each state set is linked with a label l L. Let set s = (s 1, s 2,, s T ) be a certain sequence of state, CRFs determines condition probability of sequence of state follow this function: T 1 p ( s o) expk f k ( st 1, st, o, t). Z ( o) t1 k T Call Z( o) s' exp k f k ( s' t1, s' t, o, t) t1 k standardized factor for all of possible labeled sequences. f k determines a specific function and is a linked variable that links with each f k. The purpose of machine learning of CRFs is to measure these variables. Therefore, we have two kinds of specifics f k : state specific (per-state) and transition specific(transition). f f ( perstate) k ( transition) k ( s, o, t) ( s, l) x t t k k ( o, t). ( st 1, st, t) ( st 1, l) ( st, l). Here is Kronecker-. Each per-state (2) combines label l of current state s t and a contextual predicate a binary function x k (o,t) determines the important contexts of observe o at time t. A transition specific (3) shows the dependent sequence combined with label l of previous III.1. Illustrate problems a. Goal Determine websites which contain news or not? Determine information fields which have news? Classify news? b. Solve the problem With a web site x and DOM set (document object model), nodes x 1,,x k in x. Set = y 1,,y k is labels that can be assigned to x 1,,x k with y 1,,y k are contents of nodes x 1,,x k. Examine in turn specific nodes, which include contents, to examine labels and choose label which have content, news we need. c. Appeared problems Determining websites which contain news means that when adding any websites, we have to find out whether this website contains news or not? d. Difficulties If this website contains news, search in data fields where news is contained. By had news field, determine what kind of news is? Determining had news field bases on content of news, this work relate to problem Vietnamese document processing, a tough problem. Processing websites is performed through a lot of steps which are easy to appear potential error. e. Advantages Vietnamese processing has just happened at (2) determining word-class (PosTagging), this simplifies problems. (3) Determining grammar of sentences has just stopped at step Subject - Predicate determining. Help of Vietnamese dictionaries of about words.

375 353 III.2. Some basic algorithms on combined CRFs and NLP Ideas Determining distributive probability in combined space between chain of words Sw and chain of word-class St Removing unclear word-class problem for a chain of words that can be transferred to a chain of word-class selecting such that conditional probability P(S t S w ) can combine these chains of word-class with maximum value chain of words. Process Follow Bayes formula: Determining news web pages Input: Address of web page needs to be tested and a list of sample web pages (training data set) Output: Answer for Question : This web page is news web page or not? Algorithm: web pages in sample web page set Determine field of information Input: want-to-retrieve web page Output: link news web pages and its content Algorithm: Parse text Use combine CRFs and NLP to find content of news (based on specifics of data field) o Separating Segmental to words and count number of word o Assigning labels word-class for words, count number of noun o Checking conditions of content field which has news III.3. Assigning label processing First step: Determining word-class labels for suitable word which bases on syntax rules and contexts (noun, verb,..about 48 word-classes) Second step: Starting to assign labels, find out all possible word-class labels for each remain word. Third step: Deciding to assign label results, removing unclear. Assigning label by probability method [1][2][3] With each St = t 1 t 2... t N and each S w = w 1 w 2... w N : Each P(wi w1... wi-1, t1t2...tn), suppose, appearing chance of a word, when knowing word-class label, is defined completely if we know that label. This means that P(w i w 1...w i-1,t 1 t 2...t N ) = P(w i ti). So, probability P(w 1 w 2... w N t 1 t 2...t N ) only depends on basic probabilities P(w i t i ): With probabilities P(t i t 1...t i-1 ), suppose, chance to appear of a word-class is totally clear, if we know label of word-class in a gap which has a fixed size k. This means that P(t i t 1...t i-1 ) = P(t i t i-k...t i-1 ). In general, assigned label tools often suppose k with 1(bigram) or 2 (trigram). With a library of labeled documents, parameters of this model are determined easily by Viterbi algorithm. Viterbi Algorithm With a sequence of words W 1,..., W T, word-class C 1,..., CN, probability Pr (W i C i ) and probability Bigram Pr (C i C j ), find a sequence of word-class C 1,,C T with the best suitable for a sequence of words W 1,, W T. Begin: P(S t S w ) = P(S w S t ).P(S t )/P(S w ) P(w 1 w 2... w N t 1 t 2...t N ) = P(w 1 t 1 t 2...t N ) P(w 2 w 1,t 1 t 2...t N )...P(w N w 1... w N-1, t 1 t 2...t N ) P(t 1 t 2...t N ) = P(t1)P(t 2 t 1 ) P(t 3 t 1 t 2 )... P(t N t 1...t N-1 ) P(w 1 w 2... w N t 1 t 2...t N ) = P(w 1 t 1 )P(w 2 t 2 )... P(w N t N ) for i = 1 to K do SeqScore(i,1) = Pr(C1 <Start>)* Pr(W1 Ci); BACKPTR(i,1) = 0; Repeat: for t = 2 to T do for i = 1 to K do SeqScore (i,t) = Max (SeqScore (j,t -1)* Pr (Ci Cj))* Pr (Wt Ci), với j = 1,..K; BACKPTR(i,t) = maximum value of j Determining a sequence of word-class: C(T) = i is Max of SeqScore(i,t); for i = T-1 to 1 do C(i) = BACKPTR(C(i+1),i+1);

376 354 Assigning labels based on literary style [1][2][3] Following the way that the documents represent in a specific context to determine word-class for words, and even the determining makes sure that grammar rules of words in sentence are still right. if it s in dictionary then assign it with all labels (tag) in dictionary; else assign it all possible labels (tag); j = 0; Represent through procedures: Using rules of proper nouns determining. According to determined proper nouns, continue using rules to determine 48 remain word-class labels. Assigning labels based on the combine of literary style and probability [1][2][3] Assigning label engine is a mixed system which is based on literary style and assigning label engine trigram. Processing while(j < number of labels) do o Counting P w = P(tag token) is probability from token with tag label; o Calculating P c = P(tag t 1,t 2 ), is probability of tags which appear behind the labels t 1, t 2, are suitable labels of 2 words, which is in front of token word; o Calculating P w,c = P w * P c, combines 2 above probabilities(p w and P c ); o j = j +1; end while; end for; o end while; Assigning label engine working with input data is a list of annotations, each annotation is linked with a word in document. Assigning label engine can assign a sequence which consists of 4 word-classes with probability information for each word in list, or only assigning the last result labels have the highest chance to appear. At last, we got annotations with constructed like: Algorithm Read all words in document; Assigning word-class labels for words without confusion; Using rules of proper nouns determining; According to determined proper nouns, continue using rules to determine 48 remain word-class labels; o Write to buffer; o While(buffer not empty) do Read 3 words from buffer; for each word from these 3 words do III.4. Specific of news content field News usually stay on tag [P][/P][a][/a] Length of content of news > 50 and number word (after separated segmental) > 30 Number of verbs (after assign labels PosTagger) > 2 (at least 2 sentences) Number of nouns (after assign labels PosTagger) > 10 IV. Retrieved information on new web pages with Support Vector Machine (SVM) [9][10][11] Support Vector Machine (SVM) is a classified method in machine learning, it was proposed by Vladimir and coworkers from 1970s. SVM s origine was from statistical theory, based on Structural Risk Minimisation, which worked with the idea: keeping Test Error is minimum. When it s processing, SVM will move the sample set from presenting space R n to space R d with the higher of number of dimensions. In space R d, find a optimal super plane to separate this sample set based on its classification, it means, find out distributive field of each class in space R n, and determine classification of subclass of sample. Although, this method has improved from 1970s, scientists only truly care about SVM when the first paper was placed since Application of SVM has been encounted a lot in recognize hand-writing words (Cortes and Vapnik, 1995; Scho lkopf, Burges and Vapnik, 1995; Scho lkopf, Burges and Vapnik, 1996; Burges and Scho lkopf, 1997), recognize objects (Blanz et al., 1996), recognize face in picture (Osuna, Freund and Girosi, 1997), document classification (Joachims, 1997).

377 355 Therefore, retrieved information from web with SVM is also executed by using document classification. Main idea of algorithm Check any document a i in document set a which belongs to classified A or not (information classification )? If a i A then a i is assigned label 1, else d is assigned label -1. This processing method repeats until all of documents in document set are classified and want-to-retrieve information is document assigned with the label 1. This method classifies document to get right field of information which user want, the rest is only like getting a gift on table. Algorithm [10][11] Suppose we have a set of specifics T={t 1, t 2,, t n }, each document a i is represented by a data vector x i =(w i1, w i2,, w in ), w ij R is variable of word t j in document a i. Hence, position of each data vector x i correlate with position of a point in space R n. Processing of classification will be executed on data vector x i, it s not from document d i. Therefore, we will use identically vocabularies, document, data vector, data point. Training data of SVM is labeled documents Tr={(x 1, y 1 ), (x 2, y 2 ),, (x l, y l )}(this is set of news document which was collected from news web pages), let x i be data vector, represents document d i (x i R n ), y i {+1, -1}, (x i, y i ). This means that vector x i (or document d i ) is assigned label y i. If we consider each document d i as data point in space R n, SVM will find the best geometric surface (super plane) in the space n-dimension to separate data with desire all of x + that was labeled 1, belong to positive side of super plane (f(x + )>0), all of x were labeled -1, belong to negative side of super plane (f(x + )>0). With kind of problem like SVM, a separate-data super plane is consider to be the best when distance from the nearest data point to super plane is biggest. And determining a document xtr belong to classified c or not, follow this is determining of f(x), if f(x)>0 then xc, else f(x)0 then xc. to negative side), we can find out a linear super plane to separate this data set like With wr n w T. x b 0 (4.2) is weight vector br is free coefficient so as to 1 T f xi ) sign{ w xi b} 1 y 1 i ( i i yi 1 (x,y ) Tr (4.3) Suppose that Super plane separate data and relative ties: min w i T. x i b 1 i=1,,l (4.4) T Or y w. x b1, i 1, l (4.5) i i..., The problem is how to determine w and b to get the best super plane, which distance from the nearest training data point to super plane is the farthest, with function of distance is T w. xi b d( w, b; xi ) (4.6) w T w. x b is absolute value of w T.x i +b i w is Euclid length of vector w In H3, the bold line is the best super plane, and points in box are the nearest points of super plane, called support vector. And light lines which support vector stays on were called margins. Suppose data set n x, y ),...,( x, y ), R, y { 1,1 } Tr (4.1) ( 1 1 l l Case 1 If data set can be linear Separating without trouble (all of points were assigned label 1, belongs to positive side super plane and all of points were assigned label -1, belongs x i i Suppose h(w,b)to be sum of distance from the nearest data point of class 1 to super plane and distance from the nearest data point of class -1 to super plane, therefore: h( w, b) min d( w, b; x ) mind( w, b; x ) xi, yi 1 i xi, yi 1 T T w. xi b w. xi b min min xi, yi 1 w xi, yi 1 w (4.7) 1 T T min w. xi b min w. xi b w xi, yi 1 xi, yi 1 2 w i

378 356 So, optimal super plane is a super plane with h ( w, b) 2 / w is maximum, equivalently with w is minimum. In summary, to find out the best super plane, we have to solve this optimal math 1 Min ( w) w w 2 T yi ( w. xi b) 1, 2 i 1,..., l (4.8) Case 2: Training data set Tr can be linear Separating with interference. In this case, almost points in data set are separated by linear super plane. However, a few point are interfered, mean these points have positive label but they belong to negative side of super plane and points have negative label but belong to positive side of super plane. 1 2 Min ( w, ) w C 2 T yi ( w. ( xi ) b) 1 i, i 0 l i1 i i 1,..., l i 1,..., l (4.11) To calculate directly is very difficultly and complexly. If we know Kernel function K(x i, x j ), to calculate scalar product ( x ) ( x ) in m-dimension space we needn t i j calculate directly with mapping (x i ). K x, x ) ( x ) ( x ) (4.12) ( i j i j And some functions often be used in document classification: T Linear function: K( x, x ) x x (4.13) i j Polynomial function : K(x i, x j )=(x i x j +1) d (4.14) d is a natural number, from 1 to approximate10 i j V. Results of experimental construction For combined CRFs and NLD T In this case, we replace relative tie yi ( w. xi b) 1by T relative tie yi ( w. xi b) 1 i i 1,..., l (4.9) Here, i was called slack variable, with i 0 l 1 2 Min ( w, ) w Ci 2 i1 T yi ( w. xi b) 1i, i 1,..., l (4.10) i 0 i 1,..., l C is determined parameter, define value of relative tie, the more it increase, the higher measure of violation towards experiment error. Case 3: However, not at all of training data set can be linear Separating, in this case, we will map data vector x from n-dimension space to m-dimension space (m>n), so in this m-dimension space, data set can be linear Separating. Suppose is a nonlinear mapping from space R n to space R m. : m R n R As the same time, vector x i in space R n will correlate with vector (x i ) in space R m. Our research group built a program to collect information from online Vietnamese news websites. Result: For SVM Results of experimental construction for retrieved information on web pages by SVM was executed with training data set which includes 100 pieces of news on common news web pages like The Youth, Young people, Zing, News online, Vn Express. Although both of methods can retrieve information exactly up to over 90%, combined CRFs and NLD methods bring higher result with 97% when retrieving on Zing web site. This satisfactory result got from support of NLD method, so in the future we will try to improve SVM with support of other method to raise the accuracy.

379 357 VI. Conclusion In this article the authors used and compared different methods for information extraction on the internet. The SVM and CRFs associated with natural language processing to solve to this problem. The experiment result of these methods were retrieved exactly up to 92% and 95% accuracy. The authors will research and inprove these methods with support of other method to raise the accuracy. VII. References [1] Nguyen Quang Chau, Phan Thi Tuoi, Cao Hoang Tru (2006), Assign labels Vietnamese based on the combine of literary style and probability [2] Ho Tu Bao, Luong Chi Mai, Processing Vietnamese in Information technology, page 7 [3] Dong Thi Bich Thuy, Ho Bao Quoc, Application for Natural Language Processing on search engine in Vietnamese documents [4] Trang Nhat Quang (2007), Collect information on internet to represent news on administrative web pages of city, master essay, Science and Natural University, HCMC, Vietnam [5] Nicholas Kushmerick (1997), Wrapper Induction for Information Extraction, In Proceedings of IJCAI. [6] Hongfei Qu (2001), Wrapper Induction: Construct wrappers automatically to extract information from web sources. [7] Valter Crescenzi, Giansalvatore Mecca, Paolo Merialdo (2001), RoadRunner: Towards Automatic Data Extraction from Large Web Sites, The VLDB Journal, pp [8] John Lafferty, Andrew McCallum, Fernando Pereira, Conditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data [9] Jason Weston, Support Vector Machine (and Statistical Learning Theory) Tutorial, NEC Labs America. [10] Luong Thi Minh Hong (2006), Classified document by method Support Vector Machine, master essay, Hanoi, Vietnam [11] Nguyen Linh Giang, Nguyen Manh Hien (2005), Classified Vietnamese document by SVM

380 358 Mixed-Myanmar Myanmar and English Character Recognition with Formatting Dr. Yadana Thein, Cherry Maung University of Computer Studies Yangon (U.C.S.Y) Yangon, Myanmar University of Computer Studies Yangon (U.C.S.Y) Yangon, Myanmar Abstract This paper proposed Myanmar and English typeface Character Recognition with their related format. The system converts Portable Document Format (.pdf) to Machine Editable Word Document (.doc). It includes two parts; recognition and formatting. The recognition of Myanmar and English character can be done by MICR (Myanmar Intelligent Character Recognition) which is one kind of ICR. Statistical and semantic information is used in MICR. Final decision of is made by voting system. MICR has become successful in character recognition area recent years. MICR can produce character recognition with high accuracy rate and faster speed. Table classification is used for the recognition of table format. Hough Transformation is used to detect lines in table recognition. This system can perform not only paragraph format but also text format. Paragraph format includes alignment (left, right and center). Text format includes font color, font size and bold, etc. The system use image processing and Matlab programming. Keywords: Character Recognition, MICR, Table Forma, Hough Transformation, Text Format, Paragraph Format. 1. Introduction includes Kachin, Myanmar, Rakhine and Shan, etc. Among them Myanmar language is the most commonly used. According to international language family tree, Myanmar language is a member of Sino-Tibetan language. Most of the Myanmar characters are round in shape. Myanmar characters are more complex than English alphabets and less complex than Chinese character. Software developers considered Myanmar script as a complex script. Myanmar and English character combination with formatting which is not widely popular in Myanmar computer environment will be presented. This system can produce excellent recognition and formatting rate with faster speed. 2. Myanmar and English Language Characteristic Myanmar language includes (10) digits, (33) basic characters, (12) vowels, (4) medial and other extended characters. English language consists of (26) English Capital letter and (26) small letter. The character recognition has been one of the most interesting and important fields in research world because it is a kind of communication medium between the human and computer machines. Several different methods such as artificial neural networks, multiple classifier combination, support vector machine and statistical methods have been used to recognize characters. Two main methods of Character Recognition are OCR (Optical Character Recognition) and ICR (Intelligent Character Recognition). There are more than 30,000 languages all over the world. Among them English is an international language. Even some of the Myanmar words are adopted from English language. Therefore, it is necessary to recognize both English and Myanmar characters. Myanmar language

381 359 recognition consists of the recognition of table properties such as table border color, the number of rows and columns, width and height of each cell. If the interesting region or passage is not inside the table, extract the passage outside the table. Moreover, the extraction of paragraph, row and character from both inside and outside of the table takes place. Furthermore, document format recognition is performed. The extracted characters are recognized by using MICR. MICR uses statistical and semantic information to make final decision. It produces related code in the code buffer. The UNICODE or ASCII codes are arranged in the code arrangement stage and related code of each character is assigned to the Word Document. The document format; table format, paragraph format and text format that have been recognized is applied to the Editable Word Document. The output of this system is in the form of Editable Word Document. Input (.pdf) Classification of Table from Image Fig. 1 Characteristics of English and Myanmar languages 3. Motivation and Previous Work of MICR In Myanmar character research area, only a few works for Myanmar character recognition have been studied. It is still in research because the existing works are not complete enough. The recognition of mixed Myanmar and English characters with their related format is rarely found in research environment. The recognition of characters can be done by using MICR. MICR has successfully developed in the following applications. Car license plate reader Myanmar digits recognizer Recognition of speed limited road signs Recognition of account papers and vouchers On-line Handwritten Myanmar combined words recognition system Voice production of handwritten Myanmar combined words, etc. No Extraction of passage outside the table Is it inside table? Paragraph and Row Extraction Character Extraction Yes Recognition of table format Cell Extraction 4. System Design MICR Format Extraction The input of this system is Portable Document Format which is converted to Joint Photographic Expert Group (.jpg). Image acquisition can be done by on-line or off-line technique. After image acquisition, the classification of table from the background image is carried out. If the interesting region is inside the table, table format recognition and cell extraction is performed. Table format Related code Applied Format to.doc Output (.doc) Fig. 2 Proposed system design

382 Portable Document Format (.pdf) Image acquisition is completed either by on-line or off-line technique. In on-line image acquisition is carried out by means of Tablet or PDA (personal data assistant). Off-line image acquisition is done by scanner. The input of this system is Portable Document Format (.pdf) which is highly compressed and reliable reproduction of published material. In this system, (.pdf) document of A 4 standard paper size with document format is converted to Joint Photographic Expert Group (.jpg) with PDFCreator deliver the same result as template matching, but faster. Two methods of line detection are: Cartesian parameterization y=mx + c; (1) Polar parameterization p=xcos (ø) +ysin (ø); (2).pdf document Line 1 Center Alignment Left Alignment Center Alignment Right Alignment Font Color Lines 2, 3, 4 Lines 5, 6, 7 Line 8 After line detection (a) Font size x, y coordinate TABLE x Fig. 3 Input document (.pdf) with formatting 6. Classification of Table from Image It is necessary to detect lines and intersection points in order to classify table from image. Threshold is also important for table classification because the image may include other lines. Therefore, minimum length and pixel gap are assigned as threshold value. 6.1 Hough Transformation Hough Transformation is essential to detect line segments from table. It uses evidence gathering approach. All the collinear points in line detection are stored in an accumulator array. The major advantage is that it can Angle End point y Start point (b) Fig. 4 Information of each line from the table After line detection, all the lines information are stored in an array structure. The information of each line consists of x,y coordinates of starting and ending points, angle and rho. The starting points and ending points are represented by cross illustrated in Fig.4 (a). 6.2 Recognition of table format The total number of rows can be acquired by increasing row count when the y coordinates of the first line is equal to the y coordinates of other line.

383 [ IJCSI International Journal of Computer Science Issues, Vol. 7, Issue 6, November The total number of columns can be achieved by adding one to column count each time when the x coordinates of the first line is equal to other lines in the first row. The width and height of each rectangle is measured by inches. The image size (1240x1754) is converted with the standard A4 paper size (8.27"x11.69"), then the equation becomes (150px=1inch). This system can recognize other table format such as table border color, paragraph alignment and text format for each character within each cell. The border color can be recognized by indexing rgb color of detected line. For paragraph alignment, if the paragraph is near to the left side of the cell, it is left alignment. If the paragraph is near to the right side of the cell, it is right alignment. Otherwise, the paragraph of within the cell is center alignment. Text format inside the table is the same as that of outside the table and it will be fully explained in other section. The same procedure is carried out to find the ending point of third column of the last line. B= starting point of last line; B1=B+W2; B2=B1+W3; (5) (6) Then, each cell is extracted by cropping at diagonal points shown in Fig. 5; the starting point of first line and ending point of the lower first line of first row. The same procedure is performed for the extraction of other cells. It is vital to extract each cell without border line. A A1 A2 W W1 Cropped at diagonal points H For instance, W = 324 pixel In Word Document W = 324 pixel * 10 = 3240 H = 70 pixel * 10 = 700 W 6.3 Cell Extraction Equations Derivation From pixel to inch 150 pixel = 1 inch In Word Document 1inch = 1440 The Equation for width and height 150 pixel = pixel = 10 Left alignment Center alignment Right alignment Width and Height W2 W3 B B1 B2 After cell extraction Ready for paragraph and character extraction Fig. 5 Cell extraction from table 7. Extraction of passage outside the table (1, 1) The first line contains only one starting point for all three columns. Therefore, add width of the lower first line of first row to the y coordinate of the upper first line of first row. Perform the same procedure for the third column. A= starting point of upper first line of first row; A1=A+W; (3) A2=A1+W1; (4) The last line also contains one starting points for three column. The upper first line of the last row can be achieved by subtracting the number of column from total number of lines. Next, add the width of upper first line last row to the y coordinate of the last line. (x, n) Fig. 6 Extraction of passage outside the table 3

384 362 Let (m, n) be the size of the image. The passage outside the table is extracted by cropping at the point of (1, 1) coordinate and x coordinate of the first line of the table and n. 8. Paragraph Extraction MICR. Statistic and Semantic information is acquired after preprocessing stage. MICR used statistic and semantic information. The resulting statistical or semantic information is compared with the data in the predefined database. There are three types of predefined database; (i) Basic database, (ii) Vowels database, (iii) Medial database. The final decision is made by voting system. The output of the voting system includes relevant code. This code is put into the code buffer. MICR has potential of improving efficiency in the recognition of characters. No. of char N No Line spacing >25 Line spacing >5 Fig. 7 Paragraph and row extraction Yes Statistical /Semantic Code Type W: H HC WC Loop Compare Predefined database There are two kinds of paragraph extraction. They are extraction of paragraph outside the table and extraction of paragraph inside each cell. If the line spacing is greater than 25, paragraph extraction takes place. If the line spacing is greater than 5, row extraction is carried out. Reject No Voting System Code Yes Consonant Extended Medial English 9. Character Extraction Fig. 9 MICR system flow chart c &D ; Min x, y Max x, y c & After word extraction " ; Myanmar characters Table 1: Recognition of similar patterns English Capital characters M u i C 0 O y U Fig. 8 Character extraction Character Extraction is done by using labeling method. 10. MICR (Myanmar Intelligent Character Recognition) Myanmar intelligent character recognition (MICR) is one kind of ICR. The input of MICR includes isolated characters. It is vital to perform preprocessing stage for 10.1 Statistical and Semantic Information The statistical information of typical spatial distribution of the pixel values in image can be recognized. In semantic information, some of the pixels in the image may be formed in lines or contour. Statistical and Semantic information includes the ratio of black pixel to white pixel, histogram value and pixel density, black stroke count, loop count and open position of each character.

385 363 Vertical Black Stroke Count (VC) for Myanmar and English Characters O B c Margin Left alignment Minimum x Center Maximum x Center alignment VC 3 VC 3 VC 2 Center alignment 5 Horizontal Black Stroke Count (HC) u z Loop for Myanmar and English Characters c q p B Loop 1 Loop 2 Loop 1 Loop 2 Open Direction 4 6 HC 3 HC C m 1 8 Open at Open at P-1 P-7&P-6 * HC 2 R u Loop 1 Open at P-3 Right alignment 11.2 Font size Fig. 11 Alignment The recognition of font size depends on the height of its basic consonant character. Basic characters can be divided into two groups for font size. They are one row characters (w, a, c, e, m, n, o, r, s, y, o,, u, i, v, etc.) and two rows characters (e,&,!,#, V, X, O, n, k, M, l, S, etc). For one row characters, their font size is equal to their height. For two rows character, the font size is nearly equal to half of its height. Width and Height Ratio (W: H) W x H W V 2: 1 1: 2 H W u 1: 1 H Fig. 10 Statistical and semantic information 11. Format Extraction 11.1 Alignment The alignment of the passage depends on the margin. The boundary of the passage is assumed as the margin. Minimum x is the left margin. Maximum x is the right margin. The center point of these points is (maximum x- minimum x)/2+minimum x. If the incoming paragraph is near to left margin, it is left alignment or near to right margin right alignment, otherwise center alignment. Font size Fig. 12 Font size of each character Table 2: Relation between font size and height Myanmar Characters Height of the first group 16 11/ / / / Myanmar and English Characters English Characters Height (second group Height of or Capital letter)/2 small letter / /53 55

386 G G SS IJCSI International Journal of Computer Science Issues, Vol. 7, Issue 6, November Font Color The recognition of font color can be done by indexing RGB image. After indexing, the image will contain two rgb colors; one for background color and one for foreground color. It has two options that are dither and no-dither. Dither option can cause noise and distortion of an image. Therefore, no-dither option is needed to be set. 12. Applied Format The format that has been extracted is exported to Editable Word Document with string array. The font color is stored in the string array first. Then, table format is applied to the array. Moreover, paragraph alignment is assigned to the string array. Furthermore, font size and bold are added to the array. 13. Related Code 11.4 Bold RGB image After converting RGB to indexed image Fig. 13 Font color of each character Whether the character is bold depends on the font size and the pixel count at the black stroke. The pixel count of bold character is greater than normal character. Fig. 14 Bold for each character Table 3: Pixel information at black stroke for bold and font size Font size Normal Bold 16 2px 18 3px 20 3px 22 3px 24 3px 26 3px 28 4px 36 4px 48 5px 72 8px 4px 5px 5px 5px 5px 5px 6px 6px 7px 9px In this system, the voting system produced the code number for each recognized character. These code numbers are changed into their relative Unicode or ASCII code. For example, u is the code number for u in the Unicode system. Then, these codes are arranged according to Unicode. English characters have code number. The following table shows the Unicode Sequence or ASCII code sequence for some Myanmar and English words. In Word Document, various types of font face are stored in a table. The related code is appended to the string array by indexing the related font face. The whole string array is transferred to the word document. The output of the system is Editable Word Document UNICODE In most of the world s writing systems, Unicode is allowed as a computing industry standard to represent and manipulate text. The Unicode Standard consists of a repertoire of more than 100,000 characters, a set of code charts for visual reference, an encoding methodology and set of standard character encodings, an enumeration of character properties such as upper and lower case, a set of reference data computer files, and a number of related items, such as character properties, rules for normalization, decomposition, collation, rendering and bidirectional display order. For example, udk;, u d k ; U+1000 U+102D U+102F U+1038 Table 4: Unicode Sequence of Myanmar character Myanmar Characters Unicode Sequence u n t U+1000 U+100A U+1021 s M G G S S U+103B U+103C U+103D U+103E d d D D D k k k l l l J J g m U+102B U+102C U+102D U+102E U+102F U+1030 a ; U+1031 U+1032 U+1038 U+1031 U+1032 U+1038

387 Experimental Result 13.2 ASCII Code Coding of characters have been standardize to exchange recorded data between computers efficiently. The most popular standard is ASCII (American Standard for Information Interchange). It consists of 7bits for each character. For example, c a t Table 5: ASCII codes of English alphabets English Characters A.Z a.z ASCII Sequence Fig. 15 2rows, 3columns table with format Table 7: Experimental Result of 2rows, 3columns Table 14. Limitation Table Recognition : If the line weight of table boundary is less than 1pt, Hough Transformation cannot recognize all line segments. If the line weight is greater than 3pt, Hough Transformation recognizes extra line segments. Color : If the color is very soft, the character will be misrecognized. Bold : If the font size is 24, its height is 18. When it is bold its height become 19. If the font size is 26, its height is 19. Therefore a character of font size 24 with bold, it become font size 26 with bold. Font size : In Myanmar character, the smallest font size is 16. If the font size is smaller than 16, there will be noise in it. Similar pattern : MICR misrecognize for similar pattern. Table 6: Misrecognition of some similar patterns Myanmar characters u i English characters m c 0 o y u * n " Q 2 J Cell Samples Original Document Table 8: For Character Recognition Recognition accuracy Myanmar character JPG image in pixel English character Mix-Myanmar and English 10 words 98.62% 98.68% 97.67% 50 words 95.23% 95.87% 94.98% Over % 92.42% 90.34% Output Editable Word Document Width Height Width Height Width Height

388 366 INPUT, Before Recognition 16. Conclusion Fig. 16 Final result of the proposed system In conclusion, the main contribution of this system is the recognition of table format with Myanmar character in Portable Document Format (.pdf) document. There is no research work with table recognition in Myanmar. This system produces high accuracy rate for table format such as width and height, boundary color and etc. It can also extract other format; paragraph format and text format. The recognition of characters is based on MICR. MICR can cause misrecognization for similar character pattern; c and i, m and u. Experimental result of the system mainly depends on MICR and font size. Accuracy rate of bold depend on PDFCreator. Although there is minor error, the system can produce the nearest value of the input image. Acknowledgments OUTPUT, Editable Word [4]Zar Chi Aye, Ei Ei Phyu, Yadana Thein and Myint Myint Sein, INTELLIGENT CHARACTER RECOGNITION (MICR) AND MYANMAR VOICE MIXER (MVM) SYSTEM, the 29th Asian Conference on Remote Sensing (ACRS), Colombo, Sri Lanka, [5]Swe, T. and Tin, P., Recognition and Translation of the Myanmar Printed Text Based on Hopfield Neural Network. In Proc. of 6th Asia-Pacific Symposium on Information and Telecommunication Technologies (APSITT 2005), pp , Yangon, Myanmar. [6]Chavdhuri, B. B., Pal, U. And Mitra, M., Automatic Recognition of Printed Oriya Script, Sadhana, 2002, Vol. 27, Part I [7]R. K, Rajapakse, A. R. Weerasinghe and E. K.Seneviratne, A Neural Network Based Character Recognition System for Sinhala Script, South East Asian Regonial Computer Confederation, Conference and Cyberexhibition (SEARCC 96), Bangkok, Thailand, July 4-7th [8]LI Guo-hong, SHI Peng-fei An approach to offline handwritten Chinese character recognition based on segment evaluation of adaptive duration, ISSN [9]Nafiz, A., Fatos, T.Y., Optical character recognition for cursive handwriting. IEEE Trans. On Pattern Recognition and Machine Intelligence, 24(6): [10] Dr. Yadana Thein I achieved M.Sc (Master Computer Science) in 1996 and PhD (I.T) in I am now associated professor of U.C.S.Y (University of Computer Studies, Yangon). I have written about 25 papers altogether. About 10 of them are local papers and 15 are foreign papers. My first paper is Recognition of Myanmar Handwritten Digits and Characters for ICCA conference in My current research interest is MICR (Myanmar Intelligent Character Recognition) field. Cherry Maung I am a Master Thesis Student. I got B.C.Tech in 2008 and B.C.Tech (Hons.) in I got two distinctions (English and Image Subjects) in Master course work exam. I am very grateful to my supervisor, my family and my friends for their immerse love and encouragement of the research. I want to thank all my senior friends who have done some of the previous applications of MICR. I would like to thank the reviewers and editors of IJCSI. Finally, I appreciate all the readers who spend their precious time to read this paper. References [1]Dipti Deodhare, NNR Ranga Suri, R.Amit, Preprocessing and Image Enhancement Algorithms for a Form-based Intelligent Character Recognition System, International Journal of Computer Science & Applications Vol. 2, No. 2, pp , 2005 Technomathematics Research Foundation. [2]Tay Zar Ko Ko and Dr.Yadana Thein, Converting Myanmar Portable Document Format (pdf) to Machine Editable Text with format. [3]Ei Ei Phyu, Zar Chi Aye, Ei Phyu Khaing, Yadana Thein and Myint Myint Sein, Recognition of Myanmar Handwritten Compound Words based on MICR, the 29 th Asian Conference on Remote Sensing (ACRS), Colombo, Sri Lanka, 2008

389 Personalized Online Learning with Ontological Approach 368 Bernard Renaldy Suteja 1, Suryo Guritno 2, Retantyo Wardoyo 3 and Ahmad Ashari 4 1 Department of Information Technology, Maranatha Christian University, Indonesia 2 Department of Computer Science and Electronic, Gadjah Mada University, Indonesia 3 Department of Computer Science and Electronic, Gadjah Mada University, Indonesia 4 Department of Computer Science and Electronic, Gadjah Mada University, Indonesia Abstract Learning is a cognitive activity which is different from one learner to another. Most online learning system does not participate in learner s individual aspect; ignoring the difference of specific need or personalization to the occurred cognitive experience. This research highlights the ontological approach in personalized online learning. A developed learner-role-model, which later be integrated with ontology, thus enables personalization system to guide the learner s learning process. The developed model monitors the learner s progress, in which it renews the learner s achieved knowledge and, at once, determines the next knowledge to be learned by the learner. Keywords: online learning, personalization, ontology, web semantic. 1. Introduction Personalization is the next step of the development of online learning system. Learners may be classified into certain cognitive types [1], which may vary the level of efficiency and successfulness of online learning system towards various learners. As an example of the various cognitive aspects is noticeable by solving the counting of 428 multiplied by 5 can be completed by 428 divided by 2 and then the result multiplied by 10. It is easier than directly multiplying it. This research formulates the problem related to the personalization of online learning system. An approach is aimed to the developed system based on the learner model with ontology. The system responds differently, according to the learner s character and performance and the learning substance mastered by the learner. Another important aspect is the use of Sharable Content Object Reference Model (SCORM) [2] as a standard of referred format for content development (the display of the learning object) and to do a learner model. The development of online learning system follows the learning methodology or pedagogy which constantly evolved according to the theory of distant learning (distant learning theory) by Moore [3] using the characterization of Keegan [4]. SCORM [2] consists of several technical specifications and the manual to develop learning object. SCORM is made by the initiative of Advanced Distributed Learning (ADL) for the need of Department of Defense (DoD) in the framework of web learning. SCORM functions as a facility to unite various needs and aims of different groups or organizations that work on the online learning. Specification of SCORM is divided into two parts: Content Aggregation Model, and Run-time Environment. Content Aggregation Model provides specification to develop the content based on the manual of learning object-making such as accessibility, interoperability, reusability, durability. Run-time Environment applies the mechanism of communication between Learning Management System (LMS) and learning object. SCORM is the main standard to develop online learning content and a valuable asset to LMS. Fig. 1 SCORM 1.2 component The constructed system combines the personalization with the standard of SCORM 1.2. The sample of stand-alone application to help online learning based on SCORM is exe that stands for elearning xhtml editor.

390 Fixed Model 2. Research Method Fig. 2 elearning xhtml editor The approach applied in this research method is based on the shaping of a learner model, which is developed from various scope of area such as: authoring system; user model is included by the use of semantic web, adaptive teaching system-web based, and intelligent tutoring system. Learner model defines what is noticeable related to the learner that is done by the system. This model, which is periodically shaped by system, uses data sources from learner, learner-learner interaction, lecturer or the administrator system. From the shaped learner model, a test is applicable to the different system doer i.e. the learner, teacher or content developer. 3. Model Design The model developed is called Learner Model. Learner model consists of two kinds of data; fixed and update. Fixed data is unmodified during the interaction between learner and the system. It means the data is fixed or obtained at the beginning. Update data is the output resulted from the learning progress during the interaction with system. Fixed data refers to fixed model, and the other kind of data refers to update model. Learner model is the foundation for personalization of architecture models which are developed in online learning. There are two developed architectural model in this research; online personalization and offline personalization. Online personalization monitors the learner s interaction with the system continuously at real time. It tries to adjust the content (learning material) and the navigation channel based on the learner model. Offline personalization collect the data of learner interaction towards the system, then analyze the data to recommend content alteration to content developer. Fixed model consists of five parts, each are the unity of learner s characteristic, which are unmodified during the session of online learning. They are: Personal Consists of biographical data of the learner i.e. : name, membership, Learner Activity Unit, a list of achievement, accessibility management. They are obtainable from the registration form to enroll a learning module (of a course). Personality Describe the characteristic of learner : type of personality, concentration ability (based on the average time to accomplish learning content), and the ability to interact / cooperate in teamwork or other learners and teachers. It is obtainable through Myers-Briggs test [5] Fig. 3 Personality test Cognitive Describe the ability to learn or comprehend, based on experience in interaction with the system. Obtainable through Ross and Witkin test [1]. Pedagogy Defines the learner s characteristic or learner s behavior in learning activity that refers to learner s learning model e.g.: learning style and learning approach. They are obtainable from : learning objective: a list of learning topic, learning material for learner to learn in learning module (of a course) classroom evaluation: to determine whether or not the learner take the learning evaluation control / navigation of learning module (of a course) : to determine the control type used in navigation content. Preference Preference is a group of data to customize system based on the learner s habit / like. Initially determined by system administrator. Parts of preference consists of: display format, language of content display, personalization of web-design, personalization of

391 370 command, personal notebook, voice volume, or video quality. 3.2 Update Model Update model consists of two parts: Performance It collects data related to learner performance at present in attending learning module (of a course). The data is constantly gathered to store the recent data. It is profitable from learner interaction with the system. It consists of motivation level and self-esteem in learning, learner ability to comprehend the substance of each subject, ability in general towards learning module, level of effort in the subject, and portfolio that keep all results achieved by learner in every subject (of a course). Learner s knowledge It describes the learner knowledge towards the material and the relevant competence of the enrolled subjects, i.e. : the domain of ontology which consists of all recommended learning material used in a subject information message which is applicable for the learner s active collaboration the relevant progress of material comprehension which is needed by the subject All data in the domain of ontology are gathered from learner-system interaction. 3.3 Personalization Framework Real time personalization (fig. 4) monitors the learner s real time-interaction continuously, adjusts the content (substance content) and the navigation channel of learner learning by using engine reasoning (logical reasoning) in doing the task. Further, real time personalization provides mechanism of adapting, decision and the modification. Fig. 4 real time personalization model Non-real time personalization (fig. 5) collects the data of learner-system interaction, then analyzes it using data mining tools and recommends content change to the developer (using authoring tool). Fig. 5 non-real time personalization model The implementation of this online learning system consists of well-made learner model, which enables the system to keep and to access learner data, in which to be analyzed to result different value as learner attributes. Values are obtainable by LMS from those attributes to be used further. 4. Ontology Utilization The update learner model creates a reference to enroll course topics, which is used to make a decision about the content of what should be delivered to learner. The material is managed in ontology [6], which represents a knowledge domain. Basically, ontology is a formalization of topics domain of a course, which is transformed into classes. Therefore, a connection between classes and classes attributes may appear. This model uses classes to generalize relationship between classes to form a structure of taxonomy. The subject, as it is defined in the standard of SCORM above, consists of several modules which are implemented in SCO (Shareable Content Object) or the topic of a subject. The subject must have an objective that must be achieved by learner and the process of interaction. The mentioned interaction illustrates parts of interaction between learner-system interactions, normally used to prove what has learned. Learning material is linked to the subject, objective and interactions. Those linkages are registered during learning session, which is useful to determine learner knowledge about the learned material. It enables system to predict the learner s learning progress and to notice a comprehensive material for learners. Each subject is a part of ontology which has four parameters: Correct Answer/CA (the

392 371 amount of correct answer in a test of a subject), Incorrect Answer/IA (the amount of incorrect answer in a test of a subject), Accomplished SCO Subject/AS (the amount of accomplished course), and Unaccomplished SCO Subject/US (the amount of unaccomplished course). These parameters are obtained from interaction of learner towards system, objective, and SCO data. Besides those four parameters, there is a condition in each subject within ontology, which has four values e.g. acknowledged (learner has acknowledged the concept / learning material, based on the experience. historical information of the learner), well-learned (learner does a test and obtains more than half of correct answer), learned (determining the learner s success in the material of learning which is indicated if half of the concept / learning material is accomplished), and non-learned / nonacknowledged (excluded from the other three materials). During the learner s progress, the data is renewed within the domain of ontology. Therefore, the scope of learner s knowledge can be determined in general within certain area. As an example, if a learner has accomplished modules in Table 1. It can be seen that some courses are related each other. When the learner has accomplished the modules, the amount of SCO (AS) of each material which is linked to the module will increase. Table 1: A list of certain modules and materials of related topics in learning material After accomplishing certain course topics, a questionnaire is given to the learner to test what they have learned. Table 2 shows some questions and some course topic of related subjects. System evaluates the learner s answer as it is shown in the table. For each course topic of a subject is related to the question, the learner s answer will change the amount of correct or incorrect answer. Table 2: A list of some questions and course topics of related subjects 5. Implementation 5.1 Ontology Application An applicable approach as an example is the subject of Web Wireless Programming. It defines the ontology which describes all course domains. It is connected with the interaction in each topic of the subject so there is a group of learning material in ontology as it is displayed in fig. 6. In table 3 the used data is applicable to determine the state of condition for each course topic which has been mentioned before. For each course topic of a subject, the system saves four parameters and applies the stated formula to count the state of condition of a course subject (concept s state). Table 3: The used resume data to count condition of related course topic Fig. 6 Parts of ontology which describes the example of course domain As it is shown in table 3, WML is learning material which has one correct answer and one incorrect answer. Therefore, the percentage of correct answer is not more than fifty percentages and system evaluates formula, stating the state condition of learned material to fail. Since the course topic is related to a passed SCO (AS), system evaluates formula, stating a state condition of learned material to be correct. Finally, system determines the state of course topic to be learned.