INTERNATIONAL JOURNAL COMPUTERS, …univagora.ro/m/filer_public/2012/12/07/ijcccv8n2_draft.pdfBio-Eco-Analysis for Risk Factors using GIS Software R. Serbu, S. Borza, B. Marza 252

INTERNATIONAL JOURNAL

of

COMPUTERS, COMMUNICATIONS & CONTROL

ISSN 1841-9836

ISSN-L 1841-9836

A Bimonthly JournalWith Emphasis on the Integration of Three Technologies

Year: 2013 Volume: 8 Issue: 2 (April)

This journal is a member of, and subscribes to the principles of,the Committee on Publication Ethics (COPE).

Agora University Editing House

CCC Publications

http://univagora.ro/jour/index.php/ijccc/

International Journal of Computers, Communications & Control

EDITOR IN CHIEF:Florin-Gheorghe Filip

Member of the Romanian AcademyRomanian Academy, 125, Calea Victoriei

010071 Bucharest-1, Romania, [email protected]

ASSOCIATE EDITOR IN CHIEF:Ioan Dzitac

Aurel Vlaicu University of Arad, RomaniaSt. Elena Dragoi, 2, 310330 Arad, Romania

[email protected]&

Agora University of Oradea, RomaniaPiata Tineretului, 8, 410526 Oradea, Romania

[email protected]

EXECUTIVE EDITOR:Răzvan Andonie

Central Washington University, USA400 East University Way, Ellensburg, WA 98926, USA

[email protected]

MANAGING EDITOR . . . . . . . . . . DEPUTY MANAGING EDITORMişu-Jan Manolescu Horea Oros

Agora University of Oradea, Romania University of Oradea, RomaniaPiata Tineretului, 8, 410526 Oradea St. Universitatii 1, 410087, Oradea

[email protected] [email protected]

TECHNICAL SECRETARYCristian Dziţac Emma Valeanu

R & D Agora, Romania R & D Agora, [email protected] [email protected]

EDITORIAL ADDRESS:R&D Agora Ltd. / S.C. Cercetare Dezvoltare Agora S.R.L.

Piaţa Tineretului 8, Oradea, jud. Bihor, Romania, Zip Code 410526Tel./ Fax: +40 359101032

E-mail: [email protected], [email protected], [email protected] website: http://univagora.ro/jour/index.php/ijccc/


EDITORIAL BOARD

Boldur E. BărbatLucian Blaga University of SibiuFaculty of Engineering, Department of Research5-7 Ion Raţiu St., 550012, Sibiu, [email protected]

Pierre BorneEcole Centrale de LilleCité Scientifique-BP 48Villeneuve d’Ascq Cedex, F 59651, [email protected]

Ioan BuciuUniversity of OradeaUniversitatii, 1, Oradea, [email protected]

Hariton-Nicolae CostinFaculty of Medical BioengineeringUniv. of Medicine and Pharmacy, IaşiSt. Universitatii No.16, 6600 Iaşi, [email protected]

Petre DiniCisco170 West Tasman DriveSan Jose, CA 95134, [email protected]

Antonio Di NolaDept. of Mathematics and Information SciencesUniversità degli Studi di SalernoSalerno, Via Ponte Don Melillo 84084 Fisciano,[email protected]

Ömer EgeciogluDepartment of Computer ScienceUniversity of CaliforniaSanta Barbara, CA 93106-5110, [email protected]

Constantin GaindricInstitute of Mathematics ofMoldavian Academy of SciencesKishinev, 277028, Academiei 5, [email protected]

Xiao-Shan GaoAcademy of Mathematics and System SciencesAcademia SinicaBeijing 100080, [email protected]

Kaoru HirotaHirota Lab. Dept. C.I. & S.S.Tokyo Institute of TechnologyG3-49,4259 Nagatsuta,Midori-ku,226-8502,[email protected]

George MetakidesUniversity of PatrasUniversity CampusPatras 26 504, [email protected]

Ştefan I. NitchiDepartment of Economic InformaticsBabes Bolyai University, Cluj-Napoca, RomaniaSt. T. Mihali, Nr. 58-60, 400591, [email protected]

Shimon Y. NofSchool of Industrial EngineeringPurdue UniversityGrissom Hall, West Lafayette, IN 47907, [email protected]

Stephan OlariuDepartment of Computer ScienceOld Dominion UniversityNorfolk, VA 23529-0162, [email protected]

Gheorghe PăunInstitute of Mathematicsof the Romanian AcademyBucharest, PO Box 1-764, 70700, [email protected]

Mario de J. Pérez JiménezDept. of CS and Artificial IntelligenceUniversity of Seville, Sevilla,Avda. Reina Mercedes s/n, 41012, [email protected]

Dana PetcuComputer Science DepartmentWestern University of TimisoaraV.Parvan 4, 300223 Timisoara, [email protected]

Radu Popescu-ZeletinFraunhofer Institute for OpenCommunication SystemsTechnical University Berlin, [email protected]

Imre J. RudasInstitute of Intelligent Engineering SystemsBudapest TechBudapest, Bécsi út 96/B, H-1034, [email protected]

Yong ShiResearch Center on Fictitious Economy& Data ScienceChinese Academy of SciencesBeijing 100190, [email protected] of Information Science & TechnologyUniversity of Nebraska at OmahaOmaha, NE 68182, [email protected]

Athanasios D. StyliadisAlexander Institute of TechnologyAgiou Panteleimona 24, 551 33Thessaloniki, [email protected]

Gheorghe TecuciLearning Agents CenterGeorge Mason University, USAUniversity Drive 4440, Fairfax VA [email protected]

Horia-Nicolai TeodorescuFaculty of Electronics and TelecommunicationsTechnical University “Gh. Asachi” IasiIasi, Bd. Carol I 11, 700506, [email protected]

Dan TufişResearch Institute for Artificial Intelligenceof the Romanian AcademyBucharest, “13 Septembrie” 13, 050711, [email protected]

Lotfi A. ZadehProfessor,Graduate School,Director,Berkeley Initiative in Soft Computing (BISC)Computer Science DivisionDepartment of Electrical Engineering& Computer SciencesUniversity of California Berkeley,Berkeley, CA 94720-1776, [email protected]

DATA FOR SUBSCRIBERSSupplier: Cercetare Dezvoltare Agora Srl (Research & Development Agora Ltd.)

Fiscal code: 24747462Headquarter: Oradea, Piata Tineretului Nr.8, Bihor, Romania, Zip code 410526

Bank: MILLENNIUM BANK, Bank address: Piata Unirii, str. Primariei, 2, Oradea, RomaniaIBAN Account for EURO: RO73MILB0000000000932235

SWIFT CODE (eq.BIC): MILBROBU


Short Description of IJCCC

Title of journal: International Journal of Computers, Communications & ControlAcronym: IJCCCAbbreviated Journal Title: INT J COMPUT COMMUNInternational Standard Serial Number: ISSN 1841-9836, ISSN-L 1841-9836Publisher: CCC Publications - Agora UniversityStarting year of IJCCC: 2006Founders of IJCCC: Ioan Dzitac, Florin Gheorghe Filip and Mişu-Jan ManolescuLogo:

Publication frequency: Bimonthly: Issue 1 (February); Issue 2 (April); Issue 3 (June); Issue 4(August); Issue 5 (October); Issue 6 (December).

Coverage:

• Beginning with Vol. 1 (2006), Supplementary issue: S, IJCCC is covered by Thomson Reuters -SCI Expanded and is indexed in ISI Web of Science.

• Journal Citation Reports(JCR)/Science Edition:

– Impact factor (IF): JCR2009, IF = 0.373; JCR2010, IF = 0.650; JCR2011, IF = 0.438.

• Beginning with Vol. 2 (2007), No.1, IJCCC is covered in EBSCO.

• Beginning with Vol. 3 (2008), No.1, IJCCC, is covered in Scopus.

Scope: International Journal of Computers Communications & Control is directed to the internationalcommunities of scientific researchers in computer and control from the universities, research units andindustry.

To differentiate from other similar journals, the editorial policy of IJCCC encourages the submissionof scientific papers that focus on the integration of the 3 "C" (Computing, Communication, Control).

In particular the following topics are expected to be addressed by authors:

• Integrated solutions in computer-based control and communications;

• Computational intelligence methods (with particular emphasis on fuzzy logic-based methods, ANN,evolutionary computing, collective/swarm intelligence);

• Advanced decision support systems (with particular emphasis on the usage of combined solversand/or web technologies).

Copyright c© 2006-2013 by CCC Publications


ISSN 1841-9836, ISSN-L 1841-9836, Volume 8, Issue 2, April, 2013.

Contents

IT Outsourcing - A Management-Marketing Decision

R.E. Brandabur 183

Improving Offline Handwritten Digit Recognition Using Concavity-based Features

M. Karic, G. Martinovic 191

Energy Optimization in Mobile Wireless Sensor Networks with Mobile Targets AchievingEfficient Coverage for Critical Applications

G.A. Montoya, C. Velásquez-Villada, Y. Donoso 207

Raising Energy Saving Awareness Through Educational Software

A.E. Pitic, I. Moisil, S. Dzitac 215

An Hybrid Text-Image based Authentication for Cloud Services

D.E. Popescu, A.M. Lonea 223

Enhanced Daek Block Extraction Method Performed Automatically to Determine theNumber of Clusters in Unlabeled Datasets

P. Prabhu, K. Duraiswamy 235

Bio-Eco-Analysis for Risk Factors using GIS Software

R. Serbu, S. Borza, B. Marza 252

Evaluation of an Information Assistance System Based on an Agent-Based Architecturein Transportation Domain: First Results

A. Trabelsi, H. Ezzedine 260

Author index 274

INT J COMPUT COMMUN, ISSN 1841-9836

8(2):183-190, April, 2013.

IT Outsourcing - A Management-Marketing Decision

R.E. Brandabur

Raluca Ecaterina BrandaburAcademia de Studii Economice BucurestiFacultatea de MarketingRomania, Bucureşti, Str. Mihai Eminescu nr. 13-15, sect. 1, cod 010511 4E-mail: [email protected]

Abstract: Survival of an organization in an environment increasingly aggressiverequires a more constant study of it, followed by careful planning of its activities onthe market, accordingly with the organization’s mission. In this context, targets andspecific modalities for achieving them are expressed by developing and implementingmanagement-marking strategies. Critical economic environment (recession) demandsnow thoughtful strategic movements, providing a satisfactory market share, constantcash-flow and customer loyalty. The evolution of the Information Technology (IT)allowed for several years the development of IT solutions based on Cloud Computing.The present paper describes how the decision of outsourcing IT using cloud solutionsis bounded from benefits in terms of reduced costs for the infrastructure, removes theburden of the infrastructure and networking management for the companies, offers thechance to use multi-tenant applications that are easy to be updated by the applicationdevelopers and so on.Keywords: crisis management, cloud computing, outsourcing IT, marketing-management decisions

1 Introduction

The economical cycle seems to be one of the economy’s postulates and until we learn tomanage the economic resources from a manner to linearize the economic processes, the singleavailable option is to minimize the effects, often dramatically, of the negative components: thecrisis point and the crisis itself. Due to the economic crisis, the two main objectives firmshave in this space are being able to appropriately value their portfolio by having adequate andreliable data, and cutting costs. The accent into a strong competitive environment comes onthe competitive advantage [25] and nowadays a powerful IT department can make the differencebetween failure or success. On the other hand, the complex technical nature of IT requiresexpensive equipment and therefore capital is tied up into IT department.

The survival of an organization in an environment increasingly aggressive requires a moreconstant study of it, followed by careful planning of its activities on the market in accordancewith the organization’s mission. In this context, targets and specific modalities for achievingthem are expressed by developing and implementing management-marking strategies.

Critical economic environment (recession) demands now thoughtful strategic movement pro-viding a satisfactory market share, constant cash-flow and customer loyalty.

2 Managing outsourcing strategy

Faced with the alternatives of bankruptcy or dramatically decrease of activities, most ofthe companies focus on severe costs reduction, trying to pursue in the same time the lines ofCorporate Social Responsibility (CSR) in terms of jobs preservation. One of the most effectivesolutions for costs reduction was proven to be the services outsourcing. This way, the company


184 R.E. Brandabur

can have resources at costs that allow its survival. The CSR requirements in terms of responsi-bility towards the employees are therefore respected by the fact that globally it creates new jobsinside outsourcing companies.

These kinds of practical examples such payroll, transportation, trainings or marketing servicesproved to be again practical and functional. There are various levels of outsourcing (see figure 1):high/medium and low priority. Outsourcing companies should be qualified and selected accordingto both their demonstrated effectiveness and their ability to work collaboratively. Companies cancreate real sustained value routinely and use them for far more strategic ends-to gain capabilitiesthat they don’t have in-house, or to strengthen capabilities they do have [12].

Figure 1: Strategic Outsourcing Options, adapted after [30]

Contracting third parties, specialized in areas where the firm has neither a critical strategicneed, nor special capabilities [27], enables a company to focus its efforts on its core competencies.When the outsourcing is implemented for correct reasons, it could bring some benefits in termsof cost reduction and flexibility for the organizations that help them to overcome the crisis:

* helps companies to focus on their main activities. During the financial boom period oftime, some companies grow by adding supplementary activities that weren’t in the initialportfolio. These supplementary activities need specialized personnel that couldn’t be all thetime trained inside the company and this leads to raise the company’s operating expenses.Externalizing these activities, even temporary, leads to refocusing on the core company’sactivities and increasing the business revenues.

* helps companies to optimize their costs by externalizing activities that are mandatory buthaving them inside the company is not effective (e.g. externalize the accounting activityto a specialized firm).

* saves costs related to office space by outsourcing some less important activities to thirdparties or allowing teleworking for the employees.

IT Outsourcing - A Management-Marketing Decision 185

* helps performing a risk management policy in the periods of high turnover when the incon-sistency in the employee number could increase. Outsourcing possible activities to externalcompanies could bring a certain level of stability and saves money that otherwise will bespent with new personnel training.

The key risk is the growing dependence a on an outsourcer, thus limiting company future flex-ibility. The breakdown of classical outsourcing solutions has pushed the companies to identifynew practical solutions. Long term solutions are more complex and require more than gainingcost reduction and flexibility, outsourcing strategy must provide radical change and enterprisetransformation [21].

3 Managing outsourcing IT

Early in the ’90s, various organizations are looking for improving competitive advantagesand obtaining better performance, IT being critical in enabling business development in variousdomains, IT enabled managerial innovation allowed organizations to compete more effectivelyand helped to quickly ramp up output to meet demand [8].

During the last years, the high management and the achieved high productivity involvedthe intensive use of the IT equipment. In fact, the efficient usage of the IT infrastructure wasfrom a certain point a competitive advantage. A more advanced and sophisticated equipmenthas triggered, under the astonishing rate of technological evolution, the increasing of the opera-tional costs together with the enhancing of efficiency. The attempt to keep pace with innovationswould mean permanent investments, changes and trainings, involving huge costs for the com-panies. Meanwhile, the risk of not being compatible, if not with the latest technology, at leastwith the previous one, could generate losses. Outsourcing choices represent alternate ways fororganizations to leverage available resources to increase the value of IT in meeting corporateobjectives. [20].

However we can not see IT outsourcing only as a decisional option, but rather as a pieceof managerial plans of whole business and particularly we can develop an IT strategy, whereoperational dimensions require such an approach.

Earl [4] suggests that outsourcing IT is the first option when operational performance of ITis low and is not a strength for the company, also in his "smart source" variant when businessvalue of IT is not a core of organization and operational performance of IT trough outsourcingis improved.

IT outsourcing benefits include enhanced efficiency and cost savings, infusion of cash, reducedcapital expenditure, quicker development of applications, improved services, access to new ITknowledge and technologies, and greater flexibility in IT resource management [32]. Lacityand Wilcocks [19] categorize the desired benefits of IT outsourcing in terms of six strategicfoci: financial restructuring (or cost efficiency), core competence, technology catalyst, businesstransition, business innovation and new market.

The factors that lead to success are more business oriented than anchored in technical domain.It is important to first understand the problem, then find the right operation that fits the problem.This is the case when outsource IT final results may place IT to business needs, improvingthe to management of projects change and having the appropriate balance level between themanagement expertise and technical know-how.

186 R.E. Brandabur

4 Cloud Computing - an attractive outsourcing opportunity

The evolution of the computer science technologies allowed for several years the developingof IT solutions based on Cloud Computing. Starting as a centralized process [14] IT seems toreturn at the beginnings, due to latest applications developed few years ago and tailored to thecloud technology which concentrates huge amount of IT capacity.

Into a statutory document, The National Institute of Standards and Technology [22], anagency of the U.S. Department of Commerce defines Cloud Computing as "a model for enablingubiquitous, convenient, on-demand network access to a shared pool of configurable computingresources (e.g., networks, servers, storage, applications, and services) that can be rapidly provi-sioned and released with minimal management effort or service provider interaction" [22].

Gartner defines cloud computing as a style of computing in which massively scalable IT-related capabilities are provided ’as a service’ using Internet technologies to multiple externalcustomers [10]. In a more managerial approach "cloud computing is now an emerging tech-nological phenomenon which aims to offer users complete infrastructure, platform and softwaresolutions based on a pay-as-you-go financial model and also freeing them up of managing hard-ware, software and data by moving these tasks to the Cloud providers side" [3].

The main benefits of moving toward the Cloud are: increases productivity ; reduces capitaland ongoing maintenance costs which are now transferred to the cloud provider; provides accessto the latest technologies as a basic condition for increased performance. Another importantbenefits are the ubiquitous access, global distribution and availability. Cloud applications can bedistributed anywhere around the globe and have the same highly available service levels at anylocation; measurable/metered services making possible a pay per use schema. This leads to bettercosts management and reduces the risks of paying unused resources. All of these have as resulta better consumer satisfaction. Also one of the most important advantage and characteristic ofthe Cloud is the elastic provisioning, variable capacity depending on the customer needs. TheCloud offers competitive advantage by its ease of use. [16–18]

The most important risks associated with Cloud Computing are about security and privacyof data, in terms of data storage and data transfer protection, vulnerability management andremedial, personnel and physical security, applications security, data privacy and identity man-agement, compliance requirements (e.g. disaster recovery, security standards, logs and audittrails), reliability, legal and regulatory concerns when providing cloud services. There are alsoissues related to the lack of standards for ensuring interoperability or migration between cloudproviders and not at last, inquiries around the provider’s capacity in an very young industry, interms of quality, company size, technology, communication [4, 17, 23, 29]

Related to the deployment of the cloud, NIST identifies four models: private cloud, communitycloud, public cloud and hybrid cloud [22].

Practically, public cloud vs. private cloud discussion is, at this moment, about cost of ITdepartment vs. cost of privacy and safety of organizational data. A convenient way to sharethe computation resources among many organizations is the community cloud, which, accordingto [6] can be vertical, shared by a business entity together with its partners, by a consortium orcan be shared by an IT organization to provide services to other business units.

Examples of such community clouds are: the HR-XML Consortium is the independent, non-profit, volunteer-led organization dedicated to the development and promotion of a standard suiteof XML specifications to enable e-business and the automation of human resources-related dataexchanges and Mount Sinai Hospital in Toronto is building a community cloud in conjunctionwith the Canadian government that will give 14 areas hospitals shared access to a fetal ultrasoundapplication and data storage for patient informations [7]. The main obstacle in this case is thehuge investment of consolidating cloud applications in terms of "fit" strategy between various IT


resources within cloud members. Top 10 public cloud hosting include :IBM 1, CSC 2, GoGrid 3,Joyent 4, 8x8 5, Amazon 6 [31]

When comes on business sector, public clouds seems to be a delicate subject, because asso-ciated risks, despite the fact that companies can use the public cloud with dynamic on-demandcomputing capacity much faster and without the up-front cost. For example, the software of-fered following the service model is growing at a 17 percent annual rate [1]. Under before listedconditions, choosing right a cloud computing formula became a real challenge for executives.

Externalizing IT is controversial in terms of risks, and cloud computing is even a highlycontroversial solution for outsourcing IT. Many professionals, 47% according to a 2010 study ofISACA [14] , think that benefits are fewer than the associate risks, and only 17% consider thatbenefits outweigh the risks. Same conclusion reveals a McKinsey Global Survey conducted in2010 [24]. Additional, the research revealed many barriers perceived in order to adopt the cloudtechnology from the interviewed IT personnel (462 respondents) and non IT (264 respondents)executives: evaluating and managing security or business continuity risks; managing regulatoryrisks or exposure; adapting existing business processes to cloud systems; addressing issues withmigration or interoperability with their company’s current systems or data architecture; lackof awareness or interest in cloud systems in their company; adjusting technology governanceprocesses for cloud systems (e.g. policies for control, monitoring) and developing the right set ofskills to build, manage and support cloud systems.

Opinions are also different between IT executives and non-IT executives, first suggest thatassociated risks are too high, while according to the second category: the cloud increases businessflexibility, increases the ability for IT to scale up (or shrink) to meet business needs, has lowerunit cost of IT and offer disaster recovery and business continuity [1]. According to recentstudies [24] more organizations intended to enlarge cloud computing usage within their activities(see figure 2) .

Figure 2: The Cloud Computing Adopting Model [15]

Results revealed that, from IT executives (464 respondents), more than 80% say that theircompanies are using or experimenting cloud technologies. Also 63% say that their companiesare using cloud-based applications in some aspect of day-to-day operations, and over the next12 to 18 months, deployment and piloting is expected to increase across all application types

1http://www.ibm.com2http://www.csc.com3http://www.gogrid.com4http://www.joyent.com5http://www.8x8.com6http://aws.amazon.com

188 R.E. Brandabur

explored in the survey. This study confirm one of our hypotheses: as a product, the cloudenters in its second life stage: the growth. Although, for now, it seems a viable option onlyfor the organizations that are not dealing with sensitive data or for small or medium companiesbut soon cloud computing will surpass infancy stage. However a compromise solution for thereticent companies to the outsourcing solution could be the gradual adoption of virtualization,cloud experimentation, cloud foundation, cloud exploitation and ending with the hyper cloudparadigm [15].

The latest technologies in terms of resources virtualization with direct consequences in theapparition of the cloud computing paradigm allows the relocation of the IT costs and varioususage patterns while creating new ways for individuals to consume goods and services and forentrepreneurs and enterprises to dream up viable business models. In the same time the CloudComputing makes possible the association and the collaboration between different organizationby creating virtual enterprises that can compete through their offers with the largest playersfrom their markets.

5 Future work

The author proposes to study the degree of acceptance of cloud computing solution in Ro-mania into B2B (business/organizational) market and furthermore the degree of developing ofIT outsourcing strategies.

6 Conclusions

The organizations’ infrastructure could be highly virtualized, stringing together mass quanti-ties of IT equipments into one or more easily managed logical resource pools, practically buildingthe high quality cloud computing infrastructure. The resources needed to make this happen re-quire massive investment in terms of expertise, equipment and support. Such premises conducton the strategic decision of outsourcing IT to a cloud provider as a vendor targeted for an indi-vidual market though can distinguish that provider and allow it to offer complementary servicesto the industry as well, at reasonable costs and totally customized. The decision of outsourcingIT is bounded from benefits in terms of reduced costs for the infrastructure, remove the burdenof the infrastructure and networking management for the companies, offers the chance to usemulti-tenant applications that are easy to be updated by the application developers and so on.A systematic analysis is necessary in order to support strategic marketing embraces activitiesand decisions that draw on some view of the future.

The company size and type of business make, for instant, the difference between outsourcingIT or keeping in house, but that will be not for so long. A managerial approach is needed alsoby IT service providers. They will need to own or manage the full stack of IT capabilities on amassive scale in order to provide deep expertise in delivery of services-on-demand.

Bibliography

[1] J Bughin, M. Chui, and J. Manyika.

Clouds, big data, and smart assets: Ten tech-enabled business trends to watch, August 2010.http://www.mckinsey.com/insights/mgi/in_the_news/clouds_big_data_and_smart_assets

[2] CenterBeam. [online] http://www.centerbeam.com]


[3] A. Copie. Unified model to access the cloud storage. Analele Universitatii de Vest dinTimisoara, 49(2):11–19, 2011.

[4] M. Earl. The risks of outsourcing it, April 1996. http://sloanreview.mit.edu/the-magazine/1996-spring/3732/the-risks-of-outsourcing-it/

[5] Everdream. [online] http://www.everdream.com

[6] L. Eversoll. Community cloud, 2011. http://www.lizeversoll.com/2011/01/30/community-cloud/

[7] L. Smith A health care community cloud takes shape Community cloud,2011. http://searchcio.techtarget.com/news/2240026119/ahealth- care-community-cloud-takes-shape

[8] D. FArrel, L. Mendonca, M. Nevens, J. Manyika, R. Roberts, M. Baily,T. Terwilliger, A. Webb, A. Kale, M Ramaratnam, E. Rzepniewski, N. San-thanam, and M. Cho. How it enables productivity growth, October 2002.http://www.mckinsey.com/Insights/MGI/Research/Technology_and_Innovation/How_IT_enables_productivity_growth

[9] Flickr. [online] http://www.flickr.com

[10] Gartner. Gartner says security delivered as a cloud-based service will more than triple inmany segments by 2013, 2008. [online] http://www.gartner.com/it/page.jsp?id=722307

[11] Google maps. [online] http://maps.google.com/

[12] M. Henric and B. Singh. Outsourcing can do much more than just cut costs,June 2010. [online] http://www.forbes.com/2010/06/15/outsourcing-capability-sourcing-leadership-managing-bain.html

[13] IBM. [online] http://www.ibm.com.

[14] ISACA. It control objectives for cloud computing, 2011. [online]http://www.isaca.org/Knowledge-Center/Research/Documents/ITCO_Cloud_SAMPLE_E-boo_20July2011.pdf

[15] Computing Edge. Analyzing the Differences between Cloud Computing and Virtual-ization. [online] http://computinged.com/insights/analyzing-the-differences-between-cloud-computing-and-virtualization/

[16] ISACA. Cloud computing: Business benefits with security, governanceand assurance perspectives, 2012. [online] http://www.isaca.org/Knowledge-Center/Research/ResearchDeliverables/Pages/Cloud-Computing-Business-Benefits-With-Security-Governance-and-Assurance-Perspective.aspx

[17] N. Khanapurkar. The cloud changing business ecosystem, 2011. [online]http://www.kpmg.com/IN/en/IssuesAndInsights/ThoughtLeadership/The_Cloud_Changing_the_Business_Ecosystem.pdf

[18] V. Kouyoumjian. Gis in the cloud: The new age of cloud computing and geographic informa-tion systems, June 2011. [online] http://www.esri.com/library/ebooks/gis-in-thecloud.pdf

190 R.E. Brandabur

[19] Mary Cecelia Lacity and Leslie Willcocks. Global Information Technology Outsourcing: InSearch of Business Advantage. John Wiley & Sons, Inc., New York, NY, USA, 2000.

[20] Jae-Nam Lee, Shaila M. Miranda, and Yong-Mi Kim. It outsourcing strategies: Universalis-tic, contingency, and configurational explanations of success. Info. Sys. Research, 15(2):110–131, June 2004.

[21] Jane C. Linder. Outsourcing as a strategy for driving transformation. Strategy & Leadership,Vol. 32 Is:pp.26 – 31, 2004.

[22] P. Mell and T. Grance. The NIST definition of cloud computing, 2011. [online]http://csrc.nist.gov/publications/nistpubs/800-145/SP800-145.pdf

[23] Linda Langheld Michiel Borgers, Frank Harmsen. Advances in Enterprise Engineering II,chapter Measuring the Risks of Outsourcing: Experiences from Industry, pages 197–209.Springer Berlin Heidelberg, 2009.

[24] Business Technologies Office. How it is managing new de-mands: McKinsey global survey results, November 2010. [online]http://www.mckinseyquarterly.com/How_IT_is_managing_new_demands_McKinsey_Global_Survey_results_2702

[25] Michael E. Porter. Competitive Advantage: Creating and Sustaining Superior Performance.Free Press, June 1, 1998.

[26] Automatic Data Processing. ADP. [online] http://www.adp.com

[27] J.B. Quinn. Strategic outsourcing: Leveraging knowledge capabilities, July 1999. [online]http://sloanreview.mit.edu/the-magazine/1999-summer/4041/strategic-outsourcing-leveraging-knowledge-capabilities/

[28] Amazon Web Services. Amazon Elastic Compute Cloud. [online]http://aws.amazon.com/ec2/

[29] Z. Sheng, H. Tsuji, K. Yoshida, and T. Nakatani. Preliminary Analysis for Risk Finding inOffshore Software Outsourcing from Vendors Viewpoint. Springer Berlin Heidelberg, 2009.

[30] R. Puryear. Case study, CIO Insight http://www.cioinsight.com/c/a/Case-Studies/Perspectives-Rudy- Puryear-Bain-Consulting/

[31] Business Software. Top 10 cloud hosting revealed. [online] http://www.business-software.com/offer/top-10-cloud-hosting/

[32] Peter Weill and Marianne Broadbent. Leveraging the new infrastructure: how market leaderscapitalize on information technology. Harvard Business School Press, Boston, MA, USA,1998.

[33] L. Youseff, M. Butrico, and D. Da Silva. Toward a unified ontology of cloud computing.Grid Computing Environments Workshop, 2008. GCE ’08, pages 1–10, 2011.


8(2):191-206, April, 2013.

Improving Offline Handwritten Digit Recognition UsingConcavity-based Features

M. Karic, G. Martinovic

Miran Karic, Goran MartinovicJ. J. Strossmayer University of OsijekCroatia, 31000 Osijek, Kneza Trpimira 2bE-mail: [email protected],[email protected]

Abstract:This paper examines benefits of using concavity-based structural features in recog-nition of handwritten digits. An overview of existing concavity features is presentedand a new method is introduced. These features are used as complementary fea-tures to gradient and chaincode features, both among the best performing featuresin handwritten digit recognition. Two support vector classifiers (SVCs) are chosenfor classification task as the top performers in previous works; SVC with radial basisfunction (RBF) kernel and the SVC with polynomial kernel. For reference, we alsoused the k-nearest neighbor (k-NN) classifier. Results are obtained on MNIST, USPSand DIGITS datasets. We also tested dataset independency of various feature vectorsby combining different datasets. The introduced feature extraction method gives thebest results in majority of tests.Keywords: Complementary features, concavity features, digit recognition, featureextraction, handwritten character recognition, off-line recognition.

1 Introduction

Handwritten digit recognition is an important area of optical character recognition research.Common applications are bank check processing, postal code recognition for mail sorting andrecognition of various forms for automated data entry. These applications require high recog-nition accuracy and speed. Handwritten digit recognition is also often used as a platform fortesting performance of classification algorithms.

Typical character recognition process consists of preprocessing, segmentation, feature ex-traction and classification [1]. This paper deals only with feature extraction and classification.Selection of features and classifiers is vital for performance of a recognition system. In [2] [3] itwas concluded that feature extraction is of primary importance in character recognition tasks.Even simple classifiers can give very high recognition accuracy when a well-chosen feature ex-traction method is used. A better classifier can still be used to improve the recognition accuracy.Combining classifiers is another method used to improve the accuracy; however, we will onlydiscuss single classifier recognition.

Feature extraction is a process for capturing relevant characteristics of a target object (in thiscase a digit) from an image with a fixed number of feature variables that make a feature vector.It is preferable that the size of a feature vector be as small as possible [4]. Process is sometimesskipped and the classification is performed directly on the raw image data. Numerous types offeatures for offline handwritten digit recognition exist, ranging from structural, which are basedon geometric and topological properties of a digit, to statistical, which are based on digit imagestatistical properties. Good features should maximize the between-class variance [2] [5].

Classification is a process of assigning new data to a category based on training data in knowncategories. In this paper, we use a number of human identified digit images split into trainingand test set. A classifier learns on training images and labels and produces output based on


192 M. Karic, G. Martinovic

test images. Output is then compared to test labels to evaluate the classification performance.A good classifier should be able to learn on the training data but maintain the generalizationproperty to be accurate when identifying the test set.

In [2] performance of a range of features and classifiers was tested in handwritten digitrecognition tasks. It was concluded that gradient feature [6] and chaincode feature [7], both typesof direction features, overall performed best in all tests. Furthermore, adding complementarystructural features may improve the accuracy as in [2] [8]. Support vector classifier (SVC) withradial basis function (RBF) kernel gave the highest accuracy.

This paper further investigates how complementary structural features affect overall recog-nition accuracy. Complementary features incorporate character properties that supplement pri-mary feature vector. One type of studied complementary features in [2] is concavity features. Weexpand this research with an introduced variation of concavity based complementary featuresand another variation described in [9]. Gradient and chaincode features are used as primaryfeatures. Classification is performed using the support vector classifiers and k-nearest neighbor(k-NN) classifier. New data is also used to test recognition performance.

Datasets used for the experiments are MNIST [10], USPS [11] and DIGITS [12]. The firsttwo are well known datasets in evaluation of handwritten digit recognition and classificationalgorithms [2] [13] [14], while the last one is relatively unknown [15]. All are divided into standardtraining and test sets to ensure fair comparison of different classification meethods. Transferringexpertise from one dataset to another is an unsolved problem [12] [15] so we also tested howvarious features perform on data obtained by combining different datasets.

To make this research easy to reproduce and extend, source codes in Matlab m-file formatused to extract features and links to used datasets are available online [16].

The remainder of the paper is organized as follows. Section 2 explains primary and com-plementary feature extraction methods in detail. Section 3 brings a summary of classificationmethods. Section 4 presents datasets used to test recognition performance and some previousresults on these datasets. Section 5 shows the experimental setup. Recognition results are shownand discussed in Section 6. Conclusion is given in Section 7.

2 Feature Extraction

This section presents feature extraction methods used in handwritten digit recognition experi-ments. First, the primary features are shown, namely gradient and chaincode features. Followingare complementary features based on digit image concavity. Two existing and an introduced con-cavity feature extraction methods are explained. For every method, first a brief explanation ofthe method is given, and then the setup we used in the experiments.

Source images for feature extraction are binary images. Size of all images matches the sizeof MNIST dataset images. Other datasets are converted into the format of MNIST dataset.Feature vectors are scaled to values in range 0 – 1 so that all feature variables would contributeto classification process to the same extent. Scaling is performed separately by feature extractionmethod, meaning if we use gradient and concavity feature vector, first a maximum value fmax1of gradient feature vector is found. Gradient feature vector is divided by fmax1. Concavityfeature vector is divided by its maximum value fmax2. Feature vectors are then merged.

2.1 Gradient

Gradient features, as in [1] [2] [6], are calculated by using the Sobel operator masks (figure 1)on character image to compute the gradient components on two axes. Grayscale images shouldbe used, but if source images are binary they can simply be converted into pseudogray images,

Improving Offline Handwritten Digit Recognition Using Concavity-based Features 193

as in [1] [2] [6]. Gradient strength and direction is computed from gradient components in everycharacter image pixel.

Figure 1: Sobel masks

Gradient vectors are then mapped to, most often, four or eight standard directions. Everyvector is decomposed into two components on two nearest standard directions. Figure 2 showseight standard directions and gradient vector decomposition. Character image is divided intozones by a grid, usually 4 × 4 or 5 × 5. The total sum of the component vectors is calculatedfor each standard direction in a zone. Standard direction intensities for all zones make a featurevector. Different but similar methods for gradient feature extraction can also be found in theliterature, as in [9].

Figure 2: Gradient standard directions and vector decomposition

Since the digit recognition in our experiments is performed on binary images, first an imageis converted into a pseudo-gray image. A 3 × 3 Gaussian lowpass filter with σ = 0.5 is used toblur binary image. Gradient features are then extracted using eight standard directions and a5× 5 grid. This makes a feature vector of size 5× 5× 8 = 200. A transformation on the featurevector, y = x0.5, known as Box-Cox transform [17], is carried out to make its distribution closerto the normal distribution.

2.2 Chaincode

Chaincode features, as in [2] [7] are calculated based on a character contour. Every pixelon a contour is assigned a direction code, based on its succeeding pixel’s relative position, asin figure 3. There are eight possible directions. Character image is then divided into zones bya grid, usually 4 × 4 or 5 × 5. For every zone, number of direction codes for each directionis counted, making a feature vector. Number of features can be halved by summing codes inopposite directions, as in [2].

Chaincode feature extraction is performed directly on binary digit images. 5 × 5 grid andinformation in eight directions was used to form a vector of 5 × 5 × 8 = 200 feature variables.Box-Cox transformation y = x0.5, as in [17], is used on the feature vector to make its distributioncloser to the normal distribution.

2.3 Concavity

Concavity features are used primarily as complementary features since they contain only alimited amount of information but can improve recognition accuracies of other feature extraction


(a) (b)

Figure 3: Example of a contour (a) and its chaincode values (b)

methods. They are structural features based on measurements of character concavities. Thispaper considers three different concavity feature extraction methods; two existing and one new.The first method by Favata et al. is explained in [9], we will label it conc1 for future reference.The second method by Liu et al. is shown in [2] and is labeled conc2. The introduced methodis labeled conc3. All concavity features are extracted from binary images. Below is a shortoverview of existing methods and a more extensive explanation of the new method.

Method conc1 uses a star-like operator that shoots rays in eight directions. It is observedwhat each ray hits, a character or image border. This operator is applied to every pixel on a digitimage. Unfortunately, there is no detailed explanation of the method that would allow identicalreproduction, however we believe that our reproduction should produce similar results as itutilizes the same idea. Authors themselves say that they presented one particular implementationof their philosophy and that others are possible.

Image is divided into zones by a 5×5 grid, and for every zone total number of border hits foreach direction is counted. Similar procedure is used in gradient feature extraction. To reduce thenumber of features, diagonal hits are divided among neighboring horizontal and vertical directioncounts. Total number of features is 5× 5× 4 = 100. Figure 4a shows conc1 feature extraction.

Method conc2 measures the distance from character convex hull to character pixels. Onlyhorizontal distances are calculated, both from the left and from the right. Figure 4b shows conc2feature extraction. In [2] it is used in conjunction with crossings (crs) feature extraction method.This method counts the number of transitions from black to white pixels on a binary image forevery row. Total number of features depends on image size, and for conc2 it is equal to twotimes the image height in pixels, while for crs it is equal to image height.

(a) (b)

Figure 4: conc1 (a) and conc2 (b) feature extraction

We tested by using only conc2, and by using both conc2 and crs features. Since all digitimages we used areactually 20×20 pixels centered by mass in a 28×28 pixel frame, we performedconc2 feature extraction on these 20 × 20 pixels. This gave us a feature vector of 40 variables.crs features were extracted from the whole 28× 28 image, giving additional 28 features.

The proposed concavity feature extraction method, conc3, is based on measurements of char-


acter concavity regions in a binary image. First the convex hull of a character is calculated usinga method described in [18]. By subtracting character image from the hull, concavity regions areobtained. Position of the center of mass, width, height, and the area of each outer region, i.e. onthe convex hull border, are calculated. For the inner regions, the center of mass position and thearea are calculated plus one feature variable denoting that an inner concavity exists. There are5 feature variables per outer region and 4 feature variables per inner region. Number of observedouter and inner regions must be defined in advance.

Since we use these features for handwritten digit recognition, only two inner regions arepossible and occur for number eight, thus we observe a maximum of two inner regions. Aftersome experimenting we concluded that a maximum of five outer regions give best results. Thisgives a feature vector of 2 × 4 + 5× 5 = 33 variables. Feature variables are sorted by area size,separately for inner and outer regions. Four measurements for an inner region with the largestarea are the first four variables of a feature vector. Measurements of the next inner region by areamake following four variables. Measurements of the outer regions follow, again starting from aregion with the largest area. If any of the regions does not exist, feature variables are filled withzeros, only position is filled with values 0.5, 0.5 indicating the character center. To minimizeimpact of noise and errors incurred during image retrieval, regions under a certain threshold arecut off. Figure 5 shows the conc3 feature vector variables, while figure 6 illustrates the featureextraction process.

Figure 5: conc3 feature vector. Cx and Cy are region center of mass coordinates, w and h areregion width and height.

It can be observed that conc2 features will not take into account the whole concavity of acharacter. This can be seen on figure 6c. conc3 features have the advantage of using the wholeconcavity, which is the basis for the assumption that our method will give better results.

(a) (b) (c)

Figure 6: conc3 feature extraction: (a) determining inner and outer regions, (b) extractedregions, (c) comparison of conc2 (left) and conc3 (right) features


3 Classification

Support vector machines with radial basis function (SVM-rbf) kernel and polynomial kernel(SVM-poly) are selected for classification. SVMs often give best results in digit recognition, asin [2]. k-NN classifier is included as a simple classifier for comparison with SVMs. More detailon classifiers is presented below.

3.1 Support vector classifiers

In general, support vector machines (SVMs) solve binary classification problems. Multi classclassification is accomplished by combining multiple binary SVMs [2]. For SVMs, a solution toan optimization problem is required, defined as follows [8] [13] [19]:

minω,b,ξ

{

1

2ωTω + C

l∑

i=1

ξi

}

, (1)

yi(

ωTφ (xi) + b)

≥ 1− ξi, ξi ≥ 0, i = 1, ..., l,

where (xi, yi) are training set instance-label pairs, C is the regularization parameter, w is thevector of coefficients, b a constant, ξi are parameters for handling nonseparable data (inputs)and φ maps input into higher-dimensional space. Usually the following problem is solved:

minα

{

1

2αTQα− eTα

}

, (2)

yTi α = 0, 0 ≤ αi ≤ C, i = 1, ..., l,

where e is the vector of all ones, Q is an l by l positive semidefinite matrix, Qij ≡ yiyjK (xi, xj)

are training set instance-label pairs and K (xi, xj) ≡ φ (xi)T φ (xj) is the kernel function. We

used two kernels, radial basis function (RBF) (3) and polynomial (4) kernel:

K (xi, xj) ≡ exp(

−γ ‖xi − xj‖2)

, (3)

K (xi, xj) ≡(

xTi xj)d

. (4)

To find optimal values of the variance parameter (γ) of the RBF kernel and the cost parameterC of the SVM we used a simple grid search with values 2i where i is in range -15 to 3 for γ, and-3 to 15 for C, with the minimum step value of 0.1. Similar procedure is used for the polynomialkernel of the second degree. For all SVM classification tasks we used the LIBSVM library [19].

3.2 k-NN classifier

The k-nearest neighbor algorithm (k-NN) is a method for classification based on the nearesttraining objects in the feature space. When k=1, the class of the nearest training objects becomesthe class of the test object. When k>1 the class of the test object is determined by the majorityvote of its neighbors. A weighted version of the algorithm is used in this paper. Each of theneighbors are assigned a weight equal to inverse of distance to test object. This ensures thatcloser neighbours contribute more to the decision than distant neighbours.

4 Datasets

This section presents the tested datasets. All used datasets contain grayscale images ofhandwritten digits. Since binary images are required, we converted images from all datasets tobinary form using Otsu’s global threshold method [20] [21].


4.1 MNIST

MNIST dataset is a well known dataset for handwritten digit recognition, created by YannLeCun [10]. It has 60000 training samples and 10000 test samples. All images are size 28 × 28pixels with 256 levels of gray. Actual size of digit images is 20 × 20 pixels, centered in a largerimage using center of mass. The datasets are available online, along with a list of best performingmethods and their recognition accuracies [22]. Several binarized samples from the MNIST datasetare shown in figure 7a.

4.2 USPS

USPS dataset is a US Postal handwritten digit dataset with images obtained from envelopes[11]. It has 7291 training samples and 2007 test samples. Digits are scaled to fit in 16 × 16-pixel images with 256 levels of gray. Some experimental results on USPS dataset are availableonline [23]. Several binarized samples from the USPS dataset are shown in figure 7b.

4.3 USPS-r

A modified version of USPS dataset, labeled USPS-r, is also used, since the distribution oforiginal USPS is uneven on training and test sets [24]. The modified version is created by mergingboth sets, reshuffling and then randomly dividing the full set to a new training and test set ofequal size. Each set has 4649 samples. The same dataset as in [24], available online, was used inexperiments.

4.4 DIGITS

DIGITS dataset is a less known dataset for handwritten digit recognition, described in [12].It has 1893 training samples and 1796 test samples. Digits are prepared in a similar format toUSPS dataset, with16 × 16-pixel images and 256 levels of gray. Experimental results on thisdataset can be found in [12]. Several binarized sample images from the DIGITS dataset areshown in figure 7c.

(a) (b) (c)

Figure 7: Samples from MNIST (a), USPS (b) and DIGITS (c) datasets

4.5 COMBINED

Another dataset is used, derived by combining MNIST, USPS-r and DIGITS datasets. Itstraining set actually consists of the equal number of instances from the three datasets. DIGITStraining set has the smallest number of instances therefore the equal amount of instances israndomly taken from training sets of each of the other two datasets, creating two shorter trainingsets, labelled MNIST-s and USPS-s. COMBINED dataset consists of MNIST-s, USPS-s andDIGITS training sets. 1893 instances are taken from each dataset, giving 5679 instances total.


Experiments are conducted by training on COMBINED, MNIST-s or USPS-s sets and testingon the test set of MNIST, USPS-r or DIGITS datasets. This allows us to compare recognitionaccuracy on datasets with equal training set size, when using MNIST-s and USPS-s. UsingCOMBINED set allows us to verify whether recognition accuracy can be increased by learningon instances from another dataset, and what type of features are performing best. A similarmethod is used in [15], where training on training sets of two datasets and testing on the testset of the third was performed. This gave very poor recognition accuracies, often an order ofmagnitude lower than when training on the training set of the same dataset. This might produceunreliable results therefore we decided to experiment by adding instances.

5 Experimental setup

Handwritten digit recognition performance was tested on four different datasets (MNIST, twoversions of USPS and DIGITS) using four classifiers and ten feature vectors. This gives a totalof 40 recognition accuracies for each dataset. Support vector classifiers with RBF kernel andsecond degree polynomial kernel are used for classification as described earlier. k-NN classifier isalso used for comparison. Recognition accuracy is shown for the case k = 1, i.e. 1-NN, and forthe best result for all values of k = 1, 2 . . . 11, marked k-NN, using the method described earlier.

All datasets are first converted to binary form because the features used in experimentsare extracted from binary images. Images are then converted to MNIST format. Image areacontaining a digit is scaled to size 20× 20 pixels, keeping the aspect ratio. This scaled image isthen placed on a frame of 28× 28 pixels, positioning the digit center of mass in the center of theframe.

Feature vectors are defined to contain one primary feature which is combined with comple-mentary features. Five vectors use gradient and five chaincode as primary feature extractionmethod. In addition, four out of each of the five vectors contain complementary features, conc1,conc2, conc2 with crs and conc3, while one has no complementary features. These are explainedearlier. Precise description of the feature vectors is given in table 1. Feature vectors e-grg, e-blrand e-mul exist in [2] and have the same labels.

Feature vector Features extraction methods Sizee-grg gradient 200e-grc1 gradient + conc1 300e-grc2 gradient + conc2 240e-grc2c gradient + conc2 + crs 268e-grc3 gradient + conc3 233e-blr chaincode 200e-blc1 chaincode + conc1 300e-blc2 chaincode + conc2 240e-mul chaincode + conc2 + crs 268e-blc3 chaincode + conc3 233

Table 1: Feature vectors

Following experiments on the four datasets, additional experiments with reduced trainingsets are executed. This allowed us to investigate the impact of training set size on the recogni-tion accuracy depending on the feature vector. Also we can compare recognition accuracies ondatasets when the training set sizes are equal. In addition, merging these reduced training setsallowed us to verify how the recognition accuracy is affected for different feature vectors when


we add training instances which are obtained using a different image acquisition process, in thiscase belonging to another dataset.

As defined earlier, COMBINED, MNIST-s and USPS-s sets are used for these experiments.Results when training on MNIST-s set and testing on MNIST test set, and also when trainingon USPS-s and testing on USPS-r test set are obtained. COMBINED set is used for trainingin combination with test sets of MNIST, USPS-r and DIGITS datasets. Results obtained us-ing COMBINED training set can be compared to results when using MNIST-s, USPS-s andDIGITS training sets to verify how additional instances from other datasets affect recognitionperformance.

6 Experimental results and discussion

The experimental results in form of error rates are presented in this section, followed by ourobservations and discussion. Dataset used for an experiment is given in the title of every table.In the table rows the names of the feature vectors described earlier are listed, while the columnscontain the used classifiers. Shaded results are obtained using the introduced method, whilethe best results by category are marked with bold letters. By different category we considera different primary feature, or a different classifier. More importance is given to analysis ofSVM classifier results, while the 1-NN and k-NN results are presented as an indication of SVMclassifier superiority.

6.1 Results on the individual datasets

Table 2 shows the results obtained on the MNIST dataset. The best recognition accuracy,i.e. lowest error rate (0.67%), is achieved using the e-grc3 vector and the SVM-rbf classifier. Itis also the best attained recognition accuracy in the article. Immediately behind is the accuracyof the vector e-grg (error rate 0.68%) containing only the primary feature, gradient, and nocomplementary features. Other vectors in the category achieve lower accuracies. Relationshipsamong the vectors when using the SVM-poly classifier remained approximately the same, withslightly lower accuracies, only the vector e-grc3 has an even greater advantage over the othervectors. Vectors that use chaincode primary feature achieved lower recognition accuracies. Vectore-mul has the advantage over the other vectors in its category for both SVM classifiers (errorrate 1.00 for SVM-rbf). Vector e-blc1 is second and e-blc3 third by accuracy.


Table 3 shows the results obtained on the USPS dataset. The best recognition accuracy isonce again achieved using the e-grc3 vector and the SVM-rbf classifier. Error rate of 2.39% isthe lowest among results reported in [23] using a single classifier. On top of that, we used binaryimages for training and testing which puts our method in a disadvantage. Using the SVM-polyclassifier, e-grc3 and e-grg (2.64%) have equal error rates, while other vectors have higher errorrates, i.e. decreased performance of the primary feature. Vectors with chaincode primary featuregive higher error rates with all classifiers, with vector e-blc3 performing best.

Table 4 shows the results obtained on the USPS-r dataset. Feature vector e-grc3 and SVM-rbfclassifier give the best recognition accuracy on this dataset as well (error rate 1.33%). Results aresimilar to results on USPS dataset, but with significantly higher recognition accuracies. Whenusing the SVM-poly classifier, vectors with complementary features do not increase the errorrate of the primary feature. The vectors e-grc3 and e-blc3 still give the highest recognitionaccuracies.

Table 5 shows the results obtained on the DIGITS dataset. Error rates on this dataset aresignificantly higher than on other tested datasets. Complementary features increase recognitionaccuracy in all tests. For vectors with gradient primary feature, e-grc1 gives the best performanceoverall (error rate 4.73%), e-grc2c is second (4.79%) and e-grc2 and e-grc3 with equal error ratethird (4.90%). Among vectors with chaincode primary feature, e-blc1 gives the lowest error


rate when SVM-poly classifier is used (4.90%). With SVM-rbf classifier e-mul vector achievesthe lowest error rate (5.23%). Vector e-blc3 has the lowest average error rate when both SVMclassifiers are taken into account.

Overall, according to expectations, SVM-rbf classifier gives the best recognition performance,followed by SVM-poly, while 1-NN and k-NN achieve significantly lower accuracy. The vectore-grc3 gives the highest accuracy in the majority of the tests, except for the DIGITS datasetwhere vector e-grc1 achieved higher accuracy. In addition, e-grc3 is the only vector improving onthe results of the primary feature vector e-grg in all tests, while e-grc1, e-grc2 i e-grc2c achievedhigher error rates than e-grg on several occasions. Figure 8 gives error rates on four datasetsusing the SVM-rbf classifier for vectors containing gradient primary feature. It can be seenthat concavity features improve recognition accuracy, with vector e-grc3, containing proposedconcavity features, giving best recognition performance. Vectors using the chaincode primaryfeature achieve lower recognition accuracies. Among these vectors, the best results are achievedusing the proposed vector e-blc3 and vector e-mul.

Figure 8: Error rates on four datasets for five feature vectors using SVM-rbf classifier

6.2 Results on combined datasets

The results of recognition on the MNIST test set, when training on the reduced trainingset MNIST-s, are shown in table 6. Compared to the results obtained by training on the fullMNIST training set, recognition accuracies are lower. The use of complementary features greatlyimproves the recognition accuracy, with feature vector e-grc3 performing best (error rate 1.66%using SVM-rbf). Vector e-blc3 achieves lower error rate than e-mul when reduced training set


is used. When training on the COMBINED dataset, best results are still obtained by using thevector e-grc3, as shown in table 7. Additional training images improved the recognition accuracyof e-grc3 (error rate 1.54% using SVM-rbf). Vector e-mul now has higher recognition accuracythan vector e-blc3, however it has not increased.

Table 8 shows the results obtained on the USPS-r test set, when training on the reducedtraining set USPS-s. Feature vectors e-grc3 and e-blc3 achieve best recognition accuracies intheir categories in all tests. Accuracy of the vector e-blc3 is only matched by e-mul whenusing the SVM-rbf classifier. The lowest error rate is achieved by the vector e-grc3 (2.06%).Vectors with complementary features e-grc1, e-grc2 and e-grc2c fail to decrease the error rateof the primary feature vector e-grg. Introducing additional training images overall reduced theerror rates considerably, as shown in table 9. Feature vector e-grc3 is still unmatched (errorrate 1.74%), while other vectors with complementary features again fail to decrease the errorrate of e-grg in their categories. Vector e-blc3 is unmatched among vectors using chaincodeprimary feature, decreasing error rates more then other vectors when additional training imagesare added.


Since all images of the DIGITS dataset training set are included in the COMBINED trainingset, error rates on the DIGITS dataset can be compared to error rates on the DIGITS test set,trained on the COMBINED training set, shown in table 10. Introducing additional images totraining set increased error rates on this test set. Possible explanation for this outcome is adifferent methodology used for retrieving data of DIGITS dataset and the other two datasets.Significantly higher error rates on the DIGITS dataset than on the other two datasets alsoindicate difference in the data format. Vectors with complementary features are not successful inincreasing the recognition accuracy of the gradient primary feature vector e-grg. Among vectorswith chaincode primary feature, complementary features increase the recognition accuracy, withthe vector e-mul giving the best recognition accuracy.

Best performance of the SVM-rbf classifier in combination with the vector e-grc3 was shown inthis series of experiments as well. For every vector, best recognition accuracy is given when usingthe SVM-rbf classifier. The vector e-grc3 gives the best recognition accuracy in all experimentsexcept on the DIGITS dataset. Taking into account experiments on individual and combineddatasets, the vector e-grc3 gives the highest recognition accuracy on seven out of nine differentexperimental setups. Graphical representation of the results obtained when training on reducedtraining sets is given in figure 9a and when training on combined training sets in figure 9b. It canbe concluded that the vector e-grc3 gives the best results overall. Vectors using the chaincodeprimary feature achieve lower recognition accuracies than vectors using the gradient primaryfeature. The vector e-blc3 and vector e-mul give lowest error rates in their categories.

7 Conclusion

Experimental results showed that complementary features can significantly improve recog-nition performance. The proposed concavity feature extraction method in conjunction withgradient features gave the highest recognition accuracy in majority of experiments. The methodworked well with chaincode features as well, being one out of two top performers. It also has the


(a) (b)

Figure 9: Error rates for five feature vectors using SVM-rbf classifier (a) on three datasets withequal (reduced) training set size and (b) on datasets with training on COMBINED set

lowest feature count among observed complementary features, which lowers computational costof classification. Experiments using reduced training sets showed that the proposed concavitymethod outperforms other observed approaches making it useful for applications requiring use ofa small training set. Adding training instances from another dataset reflected on the recognitionaccuracy differently for different datasets. Accuracy was increased on two datasets and decreasedon one, indicating that learning process is sensitive to small differences in image retrieval andpreprocessing. Overall, the proposed method achieved the best performance.

Acknowledgements

This work was supported by research project grant No. 165-0362980-2002 from the Ministryof Science, Education and Sports of the Republic of Croatia.

Bibliography

[1] C.-L. Liu, K. Nakashima, H. Sako, H. Fujisawa, Handwritten digit recognition: investigationof normalization and feature extraction techniques, Pattern Recognition, 37(2):265-279, 2004.

[2] C.-L. Liu, K. Nakashima, H. Sako, H. Fujisawa, Handwritten digit recognition: benchmarkingof state-of-the-art techniques, Pattern Recognition, 36(10):2271-2285, 2003.

[3] M. H. Nguyen, F. de la Torre, Optimal feature selection for support vector machines, PatternRecognition, 43(3):584-591, 2010.

[4] U. Kressel, J. SchĂźrmann, Pattern classification techniques based on function approxima-tion, Handbook of Character Recognition and Document Image Analysis, pp.49-78, 1997.

[5] B. P. Chacko, P. Babu Anto, Comparison of Statistical and Structural Features for Hand-written Numeral Recognition, Proceedings of the International Conference on ComputationalIntelligence and Multimedia Applications (ICCIMA 2007), Washington, DC (USA), pp.296-300, 2007.


[6] H. Liu, X. Ding, Handwritten Character Recognition Using Gradient Feature and QuadraticClassifier with Multiple Discrimination Schemes, Proceedings of the Eighth International Con-ference on Document Analysis and Recognition (ICDAR ’05), Washington, DC (USA), pp.19-25, 2005.

[7] O. D. Trier, A. K. Jain, T. Taxt, Feature Extraction Methods for Character Recognition - ASurvey, Pattern Recognition, 29(4):641-662, 1996.

[8] G. Vamvakas, B. Gatos, I. Pratikakis, N. Stamatopoulos, A. Roniotis, S. J. Perantonis, Hybridoff-line OCR for isolated handwritten Greek characters, Proceedings of the Fourth IASTEDInternational Conference on Signal Processing, Pattern Recognition, and Applications, Inns-bruck (Austria), pp.197-202, 2007.

[9] J. Favata, G. Srikantan, S. Srihari, Handprinted character/digit recognition using a multiplefeature/resolution philosophy, Fourth International Workshop on Frontiers in HandwritingRecognition, Taipei (Taiwan), pp. 67-70, 1994.

[10] Y. LeCun, L. Bottou, Y. Bengio, P. Haffner, Gradient-Based Learning Applied to DocumentRecognition, Proceedings of the IEEE, 86(11):2278-2324, 1998.

[11] J. J. Hull, A Database for Handwritten Text Recognition Research, Pattern Analysis andMachine Intelligence, 16(5):550-554, 1993.

[12] A. K. Seewald, Digits - A Dataset for Handwritten Digit Recognition, Austrian ResearchInstitut for Artificial Intelligence Technical Report, Vienna (Austria), 2005.

[13] C. Cortes, V. Vapnik, Support-Vector Networks, Machine Learning, 20(3):273-297, 1995.

[14] L. Van der Maaten, A New Benchmark Dataset for Handwritten Character Recognition,Tilburg University Technical Report, 2009.

[15] A. K. Seewald, On the Brittleness of Handwritten Digit Recognition Models, TechnicalReport, Seewald Solutions, Vienna (Austria), 2009.

[16] M. Karic, Concavity paper source code. [Online] Cited 2011-08-30. Available at:http://www.etfos.hr/ mkaric/conc.

[17] R. V. D. Heiden, F. C. A. Gren, The Box-Cox metric for nearest neighbor classificationimprovement, Pattern Recognition, 30(2):273-279, 1997.

[18] C. B. Barber, D. P. Dobkin, H. T. Huhdanpaa, The Quickhull Algorithm for Convex Hulls,ACM Transactions on Mathematical Software, 22(4):469-483, 1996.

[19] C.-C. Chang, C.-J. Lin, LIBSVM: a library for support vector machines, ACM Transactionson Intelligent Systems and Technology, Vol.2, No.3, pp. 27:1-27:27, 2011. Software availableat http://www.csie.ntu.edu.tw/ cjlin/libsvm.

[20] N. Otsu, A Threshold Selection Method from Gray-Level Histograms. IEEE Transactionson Systems, Man and Cybernetics, 9(1):62-66, 1979.

[21] M. R. Gupta, N. P. Jacobson, E. K. Garcia, OCR binarization and image pre-processing forsearching historical documents, Pattern Recognition, 40(2):389-397, 2007.

[22] Y. LeCun, The MNIST database of handwritten digits. [Online] Cited 2011-08-30. Availableat: http://yann.lecun.com/exdb/mnist.


[23] D. Keysers, Experimental results on the USPS database. [Online] Cited 2011-08-30. Availableat: http://www-i6. informatik.rwth-aachen.de/ keysers/Pubs/SPR2002/node10.html.

[24] C. E. Rasmussen, C. K. I. Williams, Gaussian Processes for Machine Learning, 2nd ed.,The MIT Press, 2006.


8(2):207-214, April, 2013.

Energy Optimization in Mobile Wireless Sensor Networks withMobile Targets Achieving Efficient Coverage for Critical

Applications

G.A. Montoya, C. Velásquez-Villada, Y. Donoso

Germán A. Montoya,Carlos Velásquez-Villada,Yezid DonosoUniversidad de los AndesColombia, Cra 1 Este No 19A - 40 BogotáE-mail: [email protected]@[email protected]

Abstract:The Mobile Wireless Sensor Networks (MWSN), classified within MANETS, havemultiple applications for critical situations management such as target monitoringand tracking in conflict zones, supporting urban security, critical infrastructure mon-itoring, remote locations exploration (i.e. aerospace exploration), and patients mon-itoring and care in health facilities, among others. All of these applications haverequirements of certain intelligence in the network that can be used for network’sself-configuration in order to find targets, guarantee connectivity and informationavailability until its reception.This paper proposes a MWSN architecture with an initial random distribution in aspecific work area, and a centralized management to perform autonomous decisionmaking about the movement and connectivity of the sensors. The work area presentsmobile targets with interesting events which must be covered by the mobile sensors,and thus, send the collected information through the network to any base stationavailable. Our work shows a dynamic mathematical model used to maximize targets’coverage and send its sensed information to the base stations available, while mini-mizing system’s power consumption and maximizing operation time. The heuristicalgorithm we used to construct and find a feasible solution is also shown.Keywords: MWSN, multiobjective optimization, shortest path, coverage, location,energy efficiency.

1 Introduction

Mobile Wireless Sensor Networks (MWSN) are systems with a large amount of limitationsand restrictions inherited from WSN systems but with the additional degree of liberty to move thesensor nodes which make them complex systems. Nodes’ mobility within the network enhancethe applications of these kind of systems but makes decision making more difficult, since itintroduces a new dynamic feature in the network that changes all other parameters in every stepof the movement (connectivity, bandwidth, noise and energy). However, these kind of problemscan be solved to be used in a series of critical applications with mobile scenarios like warfaresituations, where targets will be moving in time and sensors can be deployed randomly to trackthese targets and send the information back to the base; disaster recovery situations, where robotswith a central information management can move around the disaster area looking for targetslike people trapped inside structures and send this information through a direct link or usingother robots to convey the information to the base; and space exploration, where robotic roversmust move around to look for targets and communicate with satellites to send the informationback to earth.


208 G.A. Montoya, C. Velásquez-Villada, Y. Donoso

The model presented in this work assumes that every node is mobile and can act as a router,sending their own information and forwarding information from other nodes to base stations(BS). It is also assumed that the BSs don’t have sensors embedded and can’t move. Given mobiletargets, it is necessary to adjust mobile sensors positions in order to maintain connectivity andminimize energy consumption. In other words, it is possible that certain sensors must move tocover a specific target, maintaining connectivity with their neighbours to send properly collecteddata to a base station. Since all network elements can move, it is necessary to use a dynamicmathematical model that takes into account each network state in order to guarantee connectivityand minimum energy consumption.

The remainder of the paper is organized as follows: Section II states the problem and somerelated work. Section III presents the solutions’ mathematical model and computational algo-rithm, showing the methodology to implement it, and additionally, explaining several criteriato design the heuristic according to the MWSN constraints. Section IV shows some results forseveral instances with different parameters. Finally, in Section V, conclusions and future workare presented.

2 Problem statement and related work

This paper presents a difficult problem of Mobile Wireless Sensor Networks with MobileTargets (MWSN-MT), which consists of an area of interest with mobile targets, mobile sensorsand one or more base stations available inside, which can’t move and its position is known.The targets position is unknown and the sensors positions can be changed with an order froma centralized management system. The problem consists in finding all the targets inside thearea and create communication paths from them to the base stations while at the same timeminimizing the energetic cost associated to the mobile sensors sensing, movement, transmissionand reception of data. The problem can be seen as the iterative solution of multiple shortest path(SP) problems from the targets found to any base station available, multiple coverage problemswhere the mobile sensors will try to maximize the area covered looking for all the targets, intopologies were the paths may not exist or could change in future iterations.

Some related work used as basis for our research can be found in the book by Ahuja etal. [1] where the authors explain optimization for network problems and several algorithms usedto solved them. Church and ReVelle [2] state the maximal covering location problem and setthe basis for further research on the subject. Chvátal [3] states the set covering problem andsolves it through a greedy heuristic. Baldacci et al [4] combine routing and covering problemsand solve them with exact and heuristic algorithms. Powell [5] in his book states the basis ofapproximate dynamic programming (ADP) and gives models, examples and applications for itsuse. Chabini [6] presents solutions for all to one shortest path problems in discrete dynamicnetworks. Also, works on mobile robotics by Chakraborty and Sycara [7] present a dynamicapproach for a mobile robotic network where the decisions can be taken centralized or distributedwhile maintaining connectivity in the network. Zavlanos et al. [8] use a distributed hybridapproach to control mobility of the robotic network while maintaining connectivity at desiredrates. Works on mobile wireless sensor networks by Miao et al. [9] explore the problem ofdeployment and distribution of mobile sensor networks through swarm intelligence. Wang etal. [10] review mobile sensor networks challenges and applications. Finally, Cortes et al. [11]present control and coordination algorithms for mobile sensors. While the previous works serveas the basis for our research, this paper introduces the system energy consumption considerationin a distributed target searching application, a research that can lead to mobile, autonomous andefficient networks for disaster inspection and remote site exploration among other applications.

Energy Optimization in Mobile Wireless Sensor Networks with Mobile Targets AchievingEfficient Coverage for Critical Applications 209

3 Problem solution

3.1 Mathematical Formulation

The network begins with an initial graph G = (S,B, T,E,A), where S is the node setcomposed of mobile sensors, B is the set of Base Stations (nodes without mobility nor sensingcapabilities but with connectivity outside the network) and T is the targets set; E is the edgesset, which are the feasible connections between mobile sensors and between sensors and BaseStations, and A is the arcs set, which are the feasible links from targets to sensors. The graphG is converted in a new graph G′ = (S′, B′, T ′, E′, A′) after each run of the algorithm. Thisprocess is performed iteratively to reach the targets and guarantee connectivity between mobilesensors and base stations as the network elements move.

The sensors have transmitting and sensing features which follow the well known disk coveragemodel where each sensor is assumed to cover a disk centered at itself with fixed sensing andtransmitting ranges as disk radii. If a sensor cannot reach any target, it can move seeking it orserving as a relay node for communications. However, the use of these capabilities affects theenergy consumption of the network and reduces its lifetime; therefore, in the initial state, sensorshave an energy level which is reduced by movement, data transmission and reception. In eachfollowing state it is necessary to minimize the energy consumption in order to extend the lifetimeof the network and increase the probability to find targets. The system also has to maximizecoverage trying to find all targets. In figure 1, the basic structure of the optimization model isshown.

Figure 1: Optimization model structure

The mathematical model notation is shown in Table 1. A is the arcs set, representing sensorsto targets connections. E is the edges set, representing sensors to sensors or to Base Stationsconnections. rc is the maximum distance for wireless links between sensors and sensors to BaseStations and rs is the maximum distance for sensing (sensor to target). Parameters rc and rsare assumed to be constants for all nodes.

T Targets set t batik Battery level available at edge (i, k) ∈ E

S Mobile sensors set i, j Capik Capacity of edge (i, k) ∈ E

B Base stations set b distik Length of edge (i, k) ∈ E

E Edges set (i, k), i ∈ S, k ∈ S ∪B ft Data flow demand (in bps) from target t ∈ T

A Arcs set (i, t), i ∈ S, t ∈ T cik Energetic cost of edge (i, k) ∈ E

dit Cost of arc (i, t) ∈ A rc Communication radius between nodes in S ∪B

rs Sensing radius between (S) and (T )

Table 1: Mathematical model notation

Decision variables: xtik is "1" if edge (i, k) is in the path for target t to any Base Station and


"0" otherwise; and zit is "1" if target t is covered by sensor i, "0" otherwise.

minw1

∑

t∈T

∑

e∈E

cikxtik +w2

∑

(i,t)∈A

ditzit (1)

subject to:w1 + w2 = 1 (2)

Construction constraints∑

t∈T

distikxtik ≤ rc ∀(i, k) ∈ E (3)

distitzit ≤ rs ∀i ∈ S, t ∈ T (4)

Energy consumption constraints

∑

t∈T

cikxtik ≤ batik ∀(i, k) ∈ E (5)

∑

i∈S

zit = 1 ∀t ∈ T (6)

∑

t∈T

zit = 1 ∀i ∈ S (7)

Data flow and connectivity constraints

∑

t∈T

ftxtik ≤ Capik ∀(i, k) ∈ E (8)

∑

(i,t)∈A

ftzit =∑

(j,b)∈E

ftxtjb (9)

∑

j∈S

∑

t∈T

xtij −∑

j∈S

∑

t∈T

xtji = 0 ∀i ∈ S (10)

Integrality constraintsxtik ∈ {0, 1} ∀(i, k) ∈ E, t ∈ T (11)

zit ∈ {0, 1} ∀(i, t) ∈ A (12)

Constraints (3) and (4) state that the links between nodes can only exist if the distancebetween them is less than the communications coverage radius parameter (rc) and a sensor canonly cover a target if the distance from the node to the target is less than the sensing coverageradius parameter (rs). Constraint (5) states that energy consumption on a link between twonodes can’t be greater than the energy available at that link. Constraint (6) assures that atarget must be covered by at least one sensor while (7) assures that at most one sensor coversa target actively. Constraint (8) states that data flow through a link must be less or equal thanthe links capacity. Constraint (9) states that the data flow generated from the targets coveredmust arrive at the base stations. Constraint (10) is the balance constraint, stating that all theflow entering a sensor node must be equal to the flow leaving that sensor node.


3.2 Computational implementation

Figure 2 presents the solution algorithm pseudo-code. In this table the algorithms functionalcharacteristics are shown. The solution seeks to cover all targets one at a time by looking for acommunications path from a sensor covering the target to any base station in the grid. If a targetisn’t covered by a sensor, the algorithm will move the n most energetic sensors spreading themaround unexplored areas, maximizing the area coverage while looking for the target(s) missingfor a predefined number m of steps or until the target(s) is covered.

Our solution uses a greedy heuristic procedure based in minimum cost (cost is defined asa combination of distance and energy consumption). Some relaxations from the mathematicalmodel discussed before are implemented in the algorithm, those relaxations are as follows:

• Constraints (6) and (7). Each target must be covered by at most one mobile sensor.Whether there is more than one mobile sensor covering the target, it is necessary to selectthe sensor with minimum cost and greater energy. If there is no sensor covering the target,the algorithm will try to find it through sensors movement.

• Constraints (9) and (10). For a given target, it is required to find a path form the targetto any base station. Therefore, a base station is able to receive information from differenttargets, and it is possible that a base station will not be selected to receive information.Also, if the target can’t send its information to a base station, the algorithm will try tocreate a path to a base station through sensor movement.

Figure 2: Solution algorithm pseudo-code

4 Results

This section presents computational results for the heuristic algorithm described in the pre-vious section. The algorithm was coded in MATLABŽ from MathWorks. The experiments wereperformed on a CORE 2 DUO personal computer equipped with 4 GB of RAM and runningunder Microsoft Windows 7. We considered 5 random instances with parameters stated in Table2. The instances present variations in the numbers of mobile sensors, targets and base stationsand nodes and targets initial positions. Parameters like communications coverage radius (rc),sensing coverage radius (rs), initial energy level (einit), energy consumption by communications(ecomm), sensing (esens) and movement actions (emove), number of maximum steps to movelooking for a target (m) and number of mobile sensors to move in each step (n) are kept constant


0 10 20 30 40 50 60 70 80 90 1000

10

20

30

40

50

60

70

80

90

100

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

1

2

3

4

1

2

3

0 20 40 60 80 100 1200

10

20

30

40

50

60

70

80

90

100

1

2

3

4

5

6

7

8

9

10

11

12

13 14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

3233

34

35

36

37

3839

40

41

42

43

44

45

46

47

48

49

50

1

2

3

4

1

2

3

0 20 40 60 80 100 1200

20

40

60

80

100

120

1

2

3

4

5

67

8

9

10

11

12

13

14

15

16 17

18

1920

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

3637

38

39

40

41

42

43

44

45

46

47

48

49

50

1

2

3

4

1

2

3

a) b) c)

0 10 20 30 40 50 60 70 80 90 1000

10

20

30

40

50

60

70

80

90

100

1

2

3

4

5

6

7

8

9

10

11

12

13

14

1516

17

18

19

20

21

22

23

2425

26

27

28

29

30

31

3233

34

35

36

37

38

39

40

1

2

1

2

d)

0 10 20 30 40 50 60 70 80 90 1000

10

20

30

40

50

60

70

80

1

2

3

4

5

6 7

8 9

10

1112

13

14

15

16

1718

19

20

21

22

2324

25

26

27

2829

30

31

32

33

34

35

36

37

38

39

40

1

2

1

2

e)

0 50 100 150 200 2500

20

40

60

80

100

120

140

160

180

1

2

3

4

5

6

7

8 9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

2728

29

30

31

32

33

34

35

3637

38

39

40

1

2

1

2

f)

Figure 3: Graphical results for two different instances

Instance name Mobile sensors Targets Base StationsMWSN50.2.2 50 2 2MWSN50.4.3 50 4 3MWSN50.3.5 50 3 5MWSN40.2.2 40 2 2MWSN40.4.3 40 4 3

Table 2: Instances used

for all instances; these parameters values are shown in Table 3. The graphical results shown inFigure 3 are graphs representations for two different instances were a) and d) show the initialgraph and c) and f) show a feasible solution found. Restriction compliant links are painted in thefigures. Triangular nodes (N) represent the target to be covered. Circular nodes (⊙) representthe sensor nodes that can cover the targets and provide connectivity to base stations. Squarenodes (�) represent base stations nodes that can be reached to convey target information. Thepath(s) shown were found searching for the most efficient path in terms of jumps and energysaving from coverage node to base station node.

The numerical solution results for Figure 3 and the other instances in Table 2 are shown inTable 4. We performed 10 runs of the algorithm for each instance and the values for the numberof iterations, energy consumption and CPU time were averaged over those 10 runs, however inthe cases were the algorithm didn’t find a path, the results weren’t taken into account. Thealgorithm couldn’t find a first path 5% of the runs and the last path (covering all targets) 20%of the runs.

5 Conclusions and Future Work

We have presented a novel dynamic multi objective optimization approach using a greedyheuristic and a two step optimization procedure where several objectives are searched at a time


Parameter Valuerc 15[m]rs 2[m]einit 10000[mA/hr]ecomm 10[m]emove 40[m]esens 5[m]m 1[step]n 20%[nodes]

Table 3: Parameters values used for all instances

first path first path first path last path last path last pathinstance iterations energy CPU time[s] iterations energy CPU time[s]

MWSN50.2.2 120,4 18,5267 0,49452 313,2 46,4009 1,04676MWSN50.4.3 31,5 9,672 0,3354 191,8 53,8483 1,15752MWSN50.3.5 4,7 0,6747 0,21372 92,6 18,8868 0,546MWSN40.2.2 57 10,78841 0,3354 209,5 37,16683 0,74256MWSN40.4.3 52,8 20,52101 0,4368 162,8 59,17033 0,92508

Table 4: Numerical results for all instances

and where the network configuration can change at any time. The instances presented wereconstructed randomly and the algorithm was tested several times with each one of them achievingpromising results. The system was able to find a path to at least a target 95% of the times beforethe energy consumption forbids it. Also a path from all targets was achieved 80% of the times inthe instances tested, showing some baseline performance over further research can be contrastedand compared.

If the network has few mobile sensors, the movement action will waste most of the energyto find the targets, this action should be restricted and used wisely in order to enhance systemslifetime. The number of targets in the system related with the mobile sensors number willdetermine the emphasis of the algorithm, coverage or path finding. The results in this papershow that it is possible to automate the decision making in critical response time scenarios whereinformation can flow in the system and a centralized management system can take actions toensure coverage and information flow to support decision at a higher level.

Further research can be done using real life scenarios, taking into account propagation modelsfor the communications links, realistic areas with obstacles for mobility and connectivity, targetmobility models based on different applications, exact algorithms in conjunction with heuristicsand distributed computing like swarm intelligence for problem solving.

Bibliography

[1] Ahuja R. K., Magnanti T. L. and OrlinJ. B., Network Flows, Prentice-Hall, ISBN 978-0136175490, 1993.

[2] Church R. and ReVelle C., The maximal covering location problem, Papers in RegionalScience, 32(1):101-118, 1974.

[3] V. Chvátal. A greedy heuristic for the set-covering problem, Mathematics of OperationsResearch, INFORMS. 4(3):233-235, 1979.


[4] Baldacci R., Dell’Amico M., Salazar González J., The Capacitated m-Ring-Star Problem,Operations Research, 55(6):1147-1162, 2007.

[5] Powell W. B., Approximate Dynamic Programming, John Wiley & Sons, 2nd Edition, ISBN978-0-470-60445-8, 2012.

[6] Chabini I., Discrete dynamic shortest path problems in transportation applications: Complex-ity and algorithms with optimal run time, Transportation Research Record, 1645(-1):170-175,1998.

[7] Chakraborty N., Sycara K., Reconfiguration algorithms for mobile robotic networks, Roboticsand Automation (ICRA), 2010 IEEE International Conference on, pp.5484-5489, 3-7 May2010.

[8] Zavlanos M. M., Ribeiro A. and Pappas G. J., Distributed control of mobility & routingin networks of robots, Signal Processing Advances in Wireless Communications (SPAWC),2011 IEEE 12th International Workshop on, pp.236-240, 26-29 June 2011.

[9] Miao L., Qi H. and Wang F., Biologically-inspired self-deployable heterogeneous mobile sen-sor networks, Intelligent Robots and Systems 2005, (IROS 2005). 2005 IEEE/RSJ Interna-tional Conference on, pp. 2363- 2368, 2-6 Aug 2005.

[10] Wang Y.-C., Wu F.-J. and Tseng Y.-C., Mobility management algorithms and applicationsfor mobile sensor networks, Wireless Communications and Mobile Computing, 12(1):7-21,2012.

[11] Cortes J., Martinez S., Karatas T., and Bullo F., Coverage control for mobile sensing net-works, Robotics and Automation, IEEE Transactions on, 20(2):243-255, April 2004.


8(2):215-222, April, 2013.

Raising Energy Saving Awareness Through Educational Software

A.E. Pitic, I. Moisil, S. Dzitac

Alina Elena PiticLucian Blaga University of Sibiu,Faculty of Sciences, Department of Mathematics and InformaticsE-mail: [email protected]

Ioana MoisilLucian Blaga University of Sibiu,Hermann Oberth Engineering FacultyE-mail: [email protected]

Simona DzitacUniversity of Oradea, RomaniaE-mail: [email protected]

Abstract:A study that I have conducted on a sample of 395 children aged 6 to 12, from boththe urban and the rural environments, shows that an increasing number of themuse computer related technologies. Today there are an increasing number of websites that inform the user of different ways to save energy and to reduce energyconsumption because it is only natural that the modern information society would goonline to research such topics. The development of an educational application thatis focused on teaching the children about energy saving is represented on a timeline.We conducted a study in which a number of 29 children, aged between 8 and 10, usethis educational application and a questionnaire. We describe the methodology usedin the process of data gathering and then the results are interpreted. The state ofhappiness and fatigue of the child user has a great potential of influencing the way inwhich he or she uses the application, and implicitly it’s educational impact. Becausewe want to be able to reform these concepts, we will base our work on a series ofMarkov models and we will define some measures that are relevant to our goal.Keywords: energy saving, educational software, eLearning, case study.

1 Introduction

Today there are an increasing number of web sites that inform the user of different waysto save energy and to reduce energy consumption because it is only natural that the moderninformation society would go online to research such topics. The major issue with these sites isthat although scientifically correct and very complete from an informational standpoint they donot meet the user’s expectations, being much too serious and end up being considered boring.

The fast development of some technologies like computer gaming, which are very attractiveto children and teen, may present an opportunity. Indeed those who have worked with childrenand teen on a day to day basis have seen how much they love computer games. This is backedup by a number of researchers that have conducted empirical studies [7]. Thus the technologyused in computer games could be used to create educational software, raising the motivation andengagement of children and making the learning process a fun activity.

Many researchers agree to use computer games for educational purposes. In [2] we see thatcomputer games help teach children faster, the lessons being more dynamic and engaging. Thisis seen as a great alternative to the slow pace and the boredom of regular school lessons. Boyle [3]points out that computer games can lead to greater engagement and pleasure in the learning


216 A.E. Pitic, I. Moisil, S. Dzitac

process, strengthening the educational environment. Moreover there are studies that show thatcarefully selected computer games can improve thought processes [4]. In response to these studiesmany researchers have developed games for educational purposes [5], [6].

2 Initial Case Study

A study that I have conducted on a sample of 395 children aged 6 to 12, from both the ur-ban and the rural environments, shows that an increasing number of them use computer relatedtechnologies [9]. Given their exposure to these technologies it is imperative that educational ap-plications be designed in a way that takes into account the abilities, interests and the developmentdemands of children.

Figure 1 and 2 shows some interesting yet somehow expected results like the high number ofchildren having access to a computer at home, the amount of time spent by them using it or theinternet navigation preferences [9].

Figure 1: The number of children who own a PC / computer usage by theme

Figure 2: Child’s internet navigation preferences / number of hours spent in front of the PC inone day

As expected, the recreational aspect of technology usage will overcome the educational one.Possibly the most interesting question for us, "Have you ever played a computer game that hastaught you things useful for school?", was answered with YES by 72% of the children from oursample.

However, only 20% of the children have read electronic books.A pretty large number of children (18%) declare that they use computers for communication.

With regard of their age, they provided us with an interesting information on how technologiesinfluences children at small ages.

We can tell that the informational technologies is influencing children more and more by theincreasing number of hours spent by them in front of a PC (Figure 2). Only one third of those

Raising Energy Saving Awareness Through Educational Software 217

questioned spend more than two hours using a PC and of these only a handful cross into thecomputer dependent category. A probable cause for this could be the children’s relatively smallage. They are between 7 and 10 years old, primary school pupils, which means that the role oftheir teachers and parents is still leaving a powerful impression on their education. This theoryis confirmed by Figure 3 which show a connection between the child’s age and the number ofhours spent using a PC. Just as predicted, as the child’s age goes over 10 years the time spentin front of the computer rises to between three and four hours each day.

Figure 3: Chart showing child’s age and the hours per day of PC usage

Seeing such an exposure to technology it is of the utmost importance to have educationalapplications that corroborate these interests and demands for the children’s future development.

These are some of the observations that can be formulated from the initial case study data:• The time spent in from of a PC increases with the child’s age;• Multimedia applications and computer games are the preferred content for children;• The number of children using the World Wide Web is quite large;• The computer has an important (second) place in a child’s free time activities program;• Not many children have read e-books;• Educational games steadily increase in popularity among children.Here are some of our conclusions derived from our observations:• Educational applications over the World Wide Web may have a big impact;• There is little interest in text format;• The practice of using computers in the educational environment must intensify;• The process of learning by playing and the idea of learning through discovery must be made

a priority;• Educational applications should have elements that attract a child’s attention.

3 Modelling The Emotional State

We define the set of emotional states ES={pleased (P), normal (N), displeased (D)} and theset of arousal states AS={aroused (A), normal (N), sleepy (S)}.

In order to define the model we will consider:- For the emotional states (ES) we will have Pleased > Normal > Displeased;- For the arousal states (AS) we will have Aroused > Normal > Sleepy. where ">" has the

meaning of "better than".To model user emotions we use Markov chains to model the transition between different

states as shown in Figure 4 [10].The "good" transitions are the blue dotted lines (Figure 4), the red lines represent the

"bad" transitions and the black ones are "neutral" transitions. High values for P(P,P), P(N,P),


Figure 4: Modeling pleasure and arousal of the user (MES and MAS)

P(D,N), P(D,P), and low values for P(D,D), P(N,D), P(P,N) and P(P,D) will describe a goodinteraction between user and our application. The P(N,N) transition doesn’t offer informationabout emotional state.

We define a measure of "wellbeing" (user feelings during the use of an application) as dES=(α1 P(P,P)+ α2 P(N,P) + α3 P(D,N) + α4 P(D,P) - (β1 P(D,D) + β2 P(N,D) + β3 P(P,N) +β4 P(P,D)), where α = (α1, α2, α3, α4) and β = (β1, β2, β3, β4) are nonnegative real numbers,empirical determined, with at least one non-zero value.

We also define dnES=dES /(∑4

i=1 αi+∑4

i=1 βi) , a normalized measure of the "wellbeing".Greater values of dnES characterizes a good user-application intereaction.

To model user arousal, we will define the Markov model as elements of the AS set (Figure4).

Similarly MES, the blue dotted arcs represent "good" transitions, the black ones "neutral"and the red represent "bad" transitions. A good user-application interaction will have greatervalues for P(S,S), P(N,S), P(A,N) and P(A,S), and lower values for P(A,A), P(N,A), P(S,N) andP(S,A). The P(N,N) transition doesn’t offer any useful information.

We define a measure of the "fatigue stare" as dAS= α1P(A,A)+ α2P(N,A) + α3P(S,N)+ α4P(S,A) - (β1 P(S,S) + β2 P(N,S) + β3 P(A,N) + β4 P(A,S)), α = (α1, α2, α3, α4) andβ = (β1, β2, β3, β4) are nonnegative real numbers, empirical determined, with at least one non-zero value. The normalized measure of the "fatigue state" is defined as dnAS=dAS /((

∑4i=1 αi+

∑4i=1 βi)).To characterize a user session of an application we can use dnES and dnAS [10].

4 The "Energy" Educational Application

This application began it’s history in January of 2010. We wanted to create an educa-tional web application with a topic focused on the "ADAPTIVE WEB APPLICATION FORCITIZENS’ EDUCATION - TEACHING CHILDREN THE VALUE OF ELECTRICAL EN-ERGY" [1]. A succession of the most important changes suffered by the application and it’sinterface is shown in Figure 5.

We conducted a study in the 2009-2011 timeframe, on a sample of 276 children, aged 6through 12. The children have responded to a questionnaire and some partial results were usedas a starting point.

Intermediate results were used because of the ongoing study and the work on the "Energy"application. The questionnaire is made up of three open answer questions ("Do you know howto save energy?", "How is electricity produced?", "How do you save electricity?"). Methodology:• Children where given the questions during the civic education classes;


Figure 5: The versions of the "Energy" application

• The evaluator shows the questionnaire to the children in the classrom in the presence of ateacher and then distributes it to all children (5 minutes);• The children fill in the questions (a maximum of 15 minutes);• In lower classes (age group from 6 to 8) the children where given aid in reading and had

their question answered;• The questionnaires are then gathered by the evaluator/teacher.The answers where free form and they were gathered in a list of answers obtained after the

results were interpreted. The final results are given in Figure 6:

Figure 6: Did you save energy/how the energy is produced/how can you save energy?

Some of the conclusions are:• 245 of the children (88%) are aware of the importance of saving energy;• 91 children (33%) don’t know how electricity is produced;• 130 children (47%) asociate the saving of energy with turning the lights off.Based on these conclusions we chose the educational content for the first version of the

"Energy" application. A part of the results from Figure 6 where used in [1].The first design session took place in January of 2010. We chose four children (two where 8

years old and the other two 9 years old) who together with two adults formed the design team forthe application. The programming team was seldom present at the meetings. Up to the meetingin March 2010 the design team gained one 9 year old. In the beginning meetings took place ata rate of two per week, each meeting being two or three hours long.

Among the challenges encountered is the inability to capture the children attention for morethan half the meeting time. In time, by limiting meetings to two hours and by combining work


with fun the children were drawn more and more to the project. Discussions related to technologywhere always followed by examples that children could try out for themselves. The last meetingin which children where involved took place in June of 2010. The changes in the applications upto version 1.5 where of a technical nature (optimizing and bug fixing).

In October of 2011 we conducted one last study in which the children had the role of theuser. The study was comprised by the "Energy" application (Version 1.5) together with aquestionnaire.

29 children aged 8 through 10 took part in the study, all pupils of the No. 4 General Schoolin Sibiu, together with three teachers (Figure 7).

Figure 7: Study Progress

In order to determine the emotional state of the subjects we have used a variant on theAffect-Grid for Children Method [8]. Here follows a part of the results obtained from the study.

The tendency of the user to pass from an emotional state to another can be seen in Figure8. Here the initial state is given by the start of the application and the final one by exiting theapplication.

The diagrams do not take into account possible variations during the use of the application.If a child goes through a succession of emotional states like (F, N, T, N, F), the result will be 0,even if there have been variations in between.

Figure 8: Going from an emotional state to another

No observable rule can be set, users passing randomly from one emotional state to another.At first sight this result could appear strange because we were expecting the application to bea success resulting in a positive emotional state in the subjects after its use. We have stumbledupon the reason by chance and from one of the children participating in the study. The reasonsgiven by him for being sad at the end of the reserved time where that he "still wanted to play"and "didn’t want to go back to class". Having found out this we have revised the diagram, and


came up with Figure 9. In it the initial state is given by the start of the application but the finalone is given by the last page reached.

Figure 9: Going from an emotional state to another (version 2)

It is now clear that our application is interesting for the children that took part in the study:• only 9% of the have felt a drop in their happy state, a possible explination beeing that they

did not succed in finding all the usefull pages, thus not finishing the story;• 24% have felt an increase and the rest where stagnant;• None of the children where more tired after using the application;• 42% of the children felt more stimulated.Other results were obtained by interpreting the results of the questionnaire given to the

children, a part of these being shown in Figure 10.

Figure 10: What did this application teach you? / How did you save energy?

5 Conclusions

The cost of educational software for children is very high because they imply a greater effortboth in time spent and in developing abilities for working with children. One solution would be tooffer instruments that would permit software developers to test the quality of their applicationsat a lower cost, on small groups of users. The models that we have proposed permit the testingof the emotional aspect, but not the cognitive one.

Gaining children as design partners or as users is possible and useful but it presents a realchallenge. Developing educational software together with children comes with a series of inherentand specific difficulties that where surpassed by treating them as equal partners and because wehad the help of teachers trained to work with children.

We developed an application with the aid of children that where treated as design partnersand afterwards as users. A study conducted on the latest version of the application proves thatchildren like it and it has a positive influence on how they think on the subject of saving energy.


Bibliography

[1] Moisil, I., Dzitac, S., Popper, L., Pitic, A., Adaptive WEB Aplication for Citizens‘ Education- Teaching Choldren the Value of ElectricalL Energy. International Journal of Computers,Communications & Control, ISSN 1841 - 9836, VOL 5(5): 819-825, 2010.

[2] Papert, S., The Children’s Machine: Rethinking School in the Age of the Computers. NewYork: Basic Books, 1993.

[3] Boyle, T., Design for Multimedia Learning. London: Prentice Hall, 1997.

[4] Aliya, S. K., The role of computer games in the development of theoretical analysis, flexibilityand reflective thinking in children. A longitudinal study, International Journal of Psychophys-iology: 149, 2002.

[5] Virvou, M. K., Combining Software Games with Education: Evaluation of its EducationalEffectiveness. Educational Technology & Society, Vol. 8: 54-65, 2005.

[6] Conati, C., & Zhou, X., Modeling students’ emotions from cognitive appraisal in educationalgames., Proceedings of the Intelligent Tutoring Systems 2002: 994-954, 2002.

[7] Mumtaz, S., Children’s enjoyment and perception of computer use in the home and theschool, Computers and Education, Volume 36: 347-362, 2001.

[8] Widen, S. C., & Russell, J. A., Children’s Scales of Pleasure and Arousal. APS Confer-ence,Toronto: 1-8, 2001.

[9] Pitic, A., Moisil, I., Computers in School. A Case Study with children 6 to 12 Years Old. THE8TH International Scientific Conference Elearning and Software for Education Bucharest,April 26-27, 2012.

[10] Pitic, A., Moisil, I.,A novel method to characterize user sessions of educational software.3rd Word Conference on Psychology, Counselling and Guidance, Efes Kusadasi Izmir,Turkey,WCPCG 09-12 MAY 2012.


8(2):223-234, April, 2013.

An Hybrid Text-Image based Authentication for Cloud Services

D.E. Popescu, A.M. Lonea

Daniela Elena Popescu, Alina Madalina LoneaFaculty of Electrical Engineering and Information Technology, University of OradeaRomania, 410087 Oradea, 1, Armatei Romane Str.E-mail: [email protected], [email protected]

Abstract:The problem of securing access to the online information is acute today when accessto bank accounts, health records, intellectual property and business or politically sen-sitive information are made by only a few clicks, regardless of geographic location.At the same time, more and more of these accesses are made from handsets. CloudComputing is eminently suitable for addressing problems related to limited clientresources, as it offloads computation from clients and offers dynamic provisioningof compute resources. Authentication of the companys users to the cloud service ismandatory because in this way it is eliminated the attacks risks to enter into the Cloudservices. A suitable authentication is required for organizations that want to accessthe Cloud services. Our solution regards increasing security at the Security AccessPoint level of Cloud Computing and it is in fact a strong hybrid user authenticationsolution based on using image combined with text in order to avoid the weakness ofsimple user and password solution for authentication. A two factor password imagebased authentication method is proposed in this paper for cloud services. This au-thentication approach is used without additional hardware involved and presents theadvantages of utilization in terms of security and usability. Every time when the userwill be asked to provide his/her identity, a form for each image included in the photowill be listed. The user will have to remember the secret code for each image andto carefully introduce them in the forms. The global cloud access solution will bebased on our hybrid proposed text-image based solution, and will be completed bythe X.509 certificates.Keywords: authentication, multi factor password authentification, strong authen-tification, image based, cloud services, IaaS, PaaS, SaaS.

1 Introduction

As Cloud Computing (CC) model seems to be the best solution for solving the online ac-cess to services that became ubiquitous, authentication is becoming a focal point for securityprofessionals [1]. The problem of securing access to the online information is acute today whenaccess to bank accounts, health records, intellectual property and business or politically sensitiveinformation are made by only a few clicks, regardless of geographic location. At the same time,more and more of these accesses are made from handsets. This introduces security vulnerabili-ties and complications, because handsets have computational, and power limitations comparedwith traditional computers and they are constrained in terms of text input being more proneto theft than traditional computers. It is also important to point out that mobile devices inputconstraints make difficult for users to input complex passwords. Cloud Computing is eminentlysuitable for addressing problems related to limited client resources, as it offloads computationfrom clients and offers dynamic provisioning of compute resources. So, CC emerges as a new com-puting paradigm which aims to provide on-demand scalable services over the Internet via Cloudvendors to multi-tenant organizations. Enterprises are interested to move their on-premises in-frastructure into cloud computing. However they are still concerned about the security risksimplied by the act of embedding their resources within the cloud computing environment.


224 D.E. Popescu, A.M. Lonea

Authentication of the companys users to the cloud service is mandatory because in this wayit is eliminated the attacks risks to enter into the Cloud services. A suitable authentication isrequired for organizations that want to access the Cloud services. Therefore, credential man-agement, strong authentication, delegated authentication are leveraged across the cloud deliverymodels. Implementing authentication is very important, but organizations should be carefullyat the attack implications. Attacks (like: impersonation, phishing, brute force dictionary basedpassword) could occur on the credential details. Thus, authentication must be secured using thebest techniques. Decreasing the risks in the cloud environment should be the priority for theCloud providers and the organizations that adopt the cloud services. They also should select theappropriate solution in terms of cost [2].

CSA (2010) provides different recommendations for each type of the cloud services used.Thus:

• Software as a Service (SaaS) and Platform as a Service (PaaS) cloud environment provideseveral authentication options for their customers. In the case of enterprises, the Identityprovider (IdP) authenticates users and a trust relationship should be realized betweenthe organizations and the cloud services by federation. Besides the enterprises could existindividual users that will want to authenticate at the cloud services. They could do it usingthe user-centric authentication (like: Google, Yahoo ID, OpenID, Live ID etc.). Hence,those individual users will access multiple sites using a single set of credentials [2].

• Infrastructure as a Service (IaaS) cloud environment disposes by two categories of users:the enterprise IT personnel and the application users. The enterprise IT personnel are theones that develop and manage applications in the IaaS cloud model. For this type of usersthe solution that is recommended is to use a dedicated VPN with the IaaS environment, inorder to apply the existing enterprise authentication systems (e.g. Single Sign-On solutionor LDAP-based authentication) into the Cloud environment. If the VPN tunnel is notrealized for feasibility reason, then authentication assertions (SAML, WS-Federation) areapplied together with standard web encryption (SSL), which will determine the expandingof the enterprises SSO capabilities to the Cloud service. Another solution that could beimplemented in order to obtain the credentials authentication of users is to use the OpenIDoutside of the enterprise and to control the access of the users by specifying the appro-priate privileges. Furthermore, also the OATH-compliant solution (Open Authentication)could be implemented in the Cloud systems for authenticating the users. These compliantsolutions uses strong authentication [2].

Our solution regards increasing security at the Security Access Point level of CC and it is infact a strong hybrid user authentication solution based on using image combined with text inorder to avoid the weakness of simple user and password solution for authentication.

Beyond this introductory section, our paper contains other 5 sections. Section 2 points somebackground related with identity and access management in CC, section 3 emphasizes somebackground and related work concerning authentication based on image and text, section 4presents our proposed solution and section 5 contains our concluding remarks.

2 Identity and Access Management in Cloud Computing

The General Cloud Computing Architecture is composed by a massive network of "cloudservers" [3] that uses virtualization to maximize the utilization of the computing power avail-able/per server (Figure 1). According to Tianfield (2011) [4] the cloud architecture consists of

An Hybrid Text-Image based Authentication for Cloud Services 225

Cloud Platform Architecture (CPA) and Cloud Application Architecture (CAA). Clouds usersinteract with CPA and CAA using the Cloud Portal, which allows the user to select a servicefrom a service catalogue. Further, the system management will find the correct resources thatwill be allocated in the cloud by the provisioning service. The optional monitoring and meteringcomponent tracks the usage of the cloud, so the resources used can be attributed to a certainuser.

Figure 1: Cloud Computing Architecture [3]

CC offers a lot of advantages such as: it is an efficient way to store and maintain databases,being an helpful tool for business, the services offered by CC are in cloud as SaaS, cloud computingsolutions are in general less expensive than their software counterparts (pricing being offered ona per-user basis), an efficient use of CC reduce energy consumption significantly, the costumersare freed of problems related to the technological issues of installing and maintaining the IT.In [3] are identified as threats in CC:

• Abusive and Flagrant Use of CC Solution: use of stringent registration and validationprocess, improving the monitoring and coordination throughput the CC, analysing thecustomer traffic, monitoring network blocks

• Serious breach in interface and API Solution: the use of security model analysis of cloudAPIs, the implementation of a strong authentication and access controls and the evaluationof the API chain dependency

• Insider threats and attacks Solution: securing overall information, efficient compliancereporting, efficient breach notification processing

It is important to note that all the mitigation techniques proposed by [3] are related with theauthentication process.

In conclusion, Cloud providers should have established a secure access and technical solutionsfor doing it, in order to ensure that the right people access the right services. The data that willbe stored in the cloud could be accessed only by authorized users, which are specified by theprovider. The solution is to integrate the data and services accesses in the Identity and AccessManagement (IAM) infrastructure, whose requirements are [2] [5]:

1. Identity provisioning/de-provisioning

2. Authentication

3. Identity Federation

4. Access Control


Authentication and authorization aspects of cloud computing are related with various formsof identity federation and claims-based authentication that facilitate transactions between cloudentities.

In CC the servers are not accessed direct through network connections, they are accessedby the services they provide, ensuring a high degree of transparency to the cloud. Users infact access certain cloud components (request brokers) and those cloud components distributerequests to individual servers, as appropriate. This important functioning aspect of CC was useas a basis for the security components and architecture solution for CC Environments given in [6].In order to preserve the transparency character for CC, security components and services mustbe transparent and also generic - adjustable to individual users, requirements, applications, andrequired services. Further, Figure 2 was introduced and discussed, with the purpose to emphasizethe security components for CC environment. Because our paper is based on the authenticationsolution, our contribution will be on the Security Access Point (SAP) component. Hence, thesecurity components of Figure 2 are:

• The Application Access Point (AAP) Server is the service that distributes - based on typesof requests, or other parameters - cloud service requests to individual application servers.It is related and use the Services Publishing and Dispatching (SPD) Server. The SPDserver is based on the UDDI standard for discovering application services available in thecloud and it is used for publishing and discovering of cloud applications services [6].

• The Communication Access Point (CAP) is in fact the communication services provider,which is able to accept requests coming through different communications protocols [6].

• The Security Access Point (SAP) is the cloud server that provides front-end security servicesand is responsible with the authentication of users. It must be based on open standardsand applicable in an open environment [6].

• Certification Authority (CA) server provides certification services in the cloud by issuingcertificate to the client and to the SAP [6].

• The Identity Management System (IDMS) X.500 compliant directory, is another serverthat provides registration and identification services in the cloud [6].

Figure 2: Security Components and Architecture for Cloud Computing Environments [6]

We have to point out that, in order to ensure the CC security, we can implement our proposedsolution at the Client level, at the SAP level and at the AAP level, but this paper consider onlythe authentication at SAP level.


3 Background and Related Work

There are three main techniques for user authentication: knowledge based techniques, token-based techniques and techniques based on biometrics. The problem with the biometrics systemsis the difficult trade-off between impostor pass rate and false alarm rate and the fact that theyoften require specialized devices unpleasant to use. They eliminate the limitations of the humanbrain [7] Recall problems are eliminated, and so are security problems concerning users writingdown or choosing simple passwords. This eliminates nearly all of the problems with security.

Knowledge-based systems are the most frequently used for user authentication in our days.Most token-based authentication systems are also using knowledge based authentication to pre-vent impersonation through theft or loss of the token. But, the fundamental weakness ofknowledge-based authentication schemes based on recall-based authentication, is the humanlimitation to remember secure text passwords.

Text-based authentication is vulnerable to more complex attacks, such as Brute force attacksand packet sniffing. With Brute force attacks, an intruder tries to guess the users password, oruses a password hash file. Alternatively, an intruder can use easily downloaded packet sniffingtechnologies such as Ethereal (Akula). Although a random, nonsensical password offers goodsecurity, the human brain finds them almost impossible to remember.

An alternative for these knowledge-based authentication systems is to orient to the recognition-based systems [8].

The passwords have evolved from a simple dictionary or personal piece of text, to a nonsensemixture of different types of characters. This new approach of text based passwords is in conflictwith the human brains ability to remember strings. All studies made on human memory patternsshow that the brain is more adept at remembering images.

In [9] is examined the requirements of a recognition-based authentication system and proposedDeja Vu, which authenticates a user through his ability to recognize previously seen images. Theproposed authentication system is more reliable and easier to use than traditional recall-basedschemes based on user passwords or PINs and it has the advantage that it prevents users fromchoosing weak passwords and makes difficult to write down and share passwords with others.This is demonstrated by Dhamija et al (2000) in their user study where 90% of all participantssucceeded in the authentication tests using Deja Vu, and only 70% succeeded using passwordsand PINs.

Another important study in this field was made by Jackson (2006) [10], where the authorconsidered from the beginning that brute force attacks can still be a problem for an image-basedsystem, and it is important to identify the right number of combinations available that doesnot compromise the system to this type of attack, and does not overload the user with images.Furthermore, Jackson (2006)also proposed a solution against shoulder surfing, which is a gridbased image authentication system that randomize the position of different images each session(in situations where the intruder was not able to get a clear view of the image clicked, only anarea view).

The prototype designed by Jackson (2006) was used to evaluate the possibility of image-basedauthentication (IBA) method to be the main security method. He made some experimentalstudies and found that images, faces and text mixed with images seemed to offer good resultsconcerning human memory. Five experiments were made for testing security, usability, recall,methodology undertaken and whether user were able to remember passwords based on multipleimage-based interfaces. The results of these experiments indicates that the recall levels over thethree interfaces was about 90% and the best performers were obtained by the text mixed withimages (story-based interface) and images (picture based-interface). The results experimentshave been encouraging for IBA method; they showed that users were quite able to remember


Figure 3: Different image-based interfaces on security systems [10]

passwords contained on different image-based interfaces, and the human brain is able to holdsuccessfully passwords on three completely different interfaces. Furthermore, they demonstratedthat the method of combining text with images is the most effective which is the basis for ourapproach, together with the idea of randomizing the images as solution against shoulder surfing.

Nitin et al.(2008), in [11] described the new facility for authentication added to JUIT-IBAsystem which is running within the Jaypee University and Information Technology (JUIT) andis globally accessible through the website: www.juit-iba.org. Being an IBA system, it is user-friendly and it uses Kerberos protocol in order to strengthen the security during authenticationprocess. The Sign in seal advanced security feature was introduced to this system in order tomake it more secure. A sign-in seal is a secret between the computer that is setting up and IBA.Your sign-in seal is saved to your computer and it is associated with your computer. if you login from multiple computers, you will need to create a separate sign-in seal for each one. It isconvenient to instantly recognize a genuine IBA sign-in page that ensure you that you are noton a page created by hackers attempting to steal your IBA ID and password The seal can becustomized by creating a text seal or by uploading an image. From the security point of view itis important that even if the hacker knows or guesses the ID on your personal information, hecannot use it to discover your sign in seal.

Even Yahoo has implemented the sign in seal method, with a seal that can be a text or anuploaded image (Figure 4) and it must be used in combination with our proposed method inorder to increase the security. This will be subject for our future work.

Figure 4: Sign in seal at yahoo


Newman et al. (2005) presents and analyses in [12] a user authentication techniques us-ing images that can be used in local or in remote authentication. The system consists of anauthentication server (AS), and an authentication user agent (AUA) and it requires that theuser have assigned a subset of images (as passwords) from a larger set. The set of all imagesused by IBA system, named image set, contains images that are distinctive to the human eye,they are not easily describable and they differ in structure. The AS has the authenticationdatabase of images and associations of users with their individual image sets. It is part of thetrusted computing base, being never compromised. When the authentication is made remote,the channel is encrypted using Diffie-Hellman. For the attack scenarios, they considered fourlocations of vulnerabilities: information stored on the AS, information sent between AS and theAUA, the output of the AUA and the input of the AUA. For security analysis purposes, theyconsidered the situations: keystroke logging: AUA Input, shoulder surfing: AUA output logging,TEMPEST Attack: AUA Output, Brute force attack, Frequency correlation attack: presentationsets, leaking image set size. For storing the individual image set for each individual user, onlythe indices into the image space will be stored. If the encryption is required, a good proposedsolution is to send the images in clear over the channel with the permutation encrypted. Thehidden permutation is applied by AUA to the images in order to display them, record the usersselections and sent these back to the AS. Their study was important for our proposed solutionand we apply our solution within the AS and AUA components, which are called in this paperas SAP (Security Access Point) authentication solution.

In the past few years, an important research was made regarding the usability and securityof challenge questions, that are commonly used as a backup when users forgot their main"authentication secret. Most challenge questions rely on a user’s knowledge of their early life,something static over time. This kind of information can be discovered by a determined attacker.

So, the standard mechanism based on textual questions can be replaced by the challengeprotocol developed by Renaud et al (2010) in [14]; it uses a set of pictorial elements to promptanswers. The prompts solicit associative memories and serve as a stronger cue to aid the recall.All the pictures serve as an additional recall aid, while the use of an indirect question (the directanswer is not in a database or a public source) helps to reduce the exposure of the user to targetedobservation attacks [15]. By using this more usable picture-based system it is maintained thesame level of security as traditional questions as long as multiple questions are used in serialorder.

Another approach for authentication was proposed by Micallef, et al. (2009) and uses anAuthentication Avatar which represents the identity, including personality, of a fictional personthat is generated almost randomly from a minimal user input [16]. An Avatar Profile (AP)contains information about the avatar, and a subset of the AP information is used by the userto respond to challenge questions regarding the avatar. In this way the security is improvedsince, unlike the users own information, the avatar information is not as easily determined byan attacker. Since such fiction information is likely to be more challenging for a user to recall(than their own, personal information), the proposed approach uses techniques such as repeatedexposure to graphical imagery (related to the avatar) of users at every login in order to improvethe memory association. Such images can be associated with the avatar itself, and also withelements of the AP (a picture of the Avatars pet). This recovering password solution was usedas basis in our authentication solution, for increasing the security level in authentication.

The proposed solutions from [14] and [16] can be integrated into the global authenticationprocess. Our proposed approach share with them the idea of combining text with images andthe idea of using a randomize process in the authentication procedure.

Another work in context of IBA was realized by Confident Technologies. They providesimage-based, multifactor authentication solutions for enterprise companies, websites, web and


mobile applications, and mobile devices [13]. It encrypts one-time authentication codes within animage-based challenge, being easy to use and highly secure. Users simply identify which picturesmatch their previously-chosen, secret categories to authenticate. Image-based authenticationsolves the traditional trade-off between security and usability by providing strong authenticationthat is easy for people to use. As users simply tap a few pictures to authenticate, it is ideallysuited for use on mobile devices.

Confident technology can be used as a standalone multifactor authentication solution, or asan additional layer of authentication.

4 The proposed authentication solution

The system consists of an authentication service server (AS), and an authentication useragent (AUA) and it requires that the user have assigned a subset of images (as passwords) froma larger set. The set of all images used by the Image Based Authentication (IBA) system, namedimage set, contains images that are distinctive to the human eye, they are not easily describableand they differ in structure. The AS has access to the authentication database of images andassociations of users with their individual image sets. The current solution for authenticating inCC is given in Figure 5 and concerns the access of the user to the Security Access Point (SAP).

Figure 5: The current CC SAP authentification Solution

In order to perform stronger authentication for accessing the cloud services, we propose touse an authentication scheme (Figure 6) that perform the following steps:

1. Make a text-based authentication based on the user ID and a text password to have accessto the cloud services

2. Make an hybrid text-image based authentication that uses our own proposed solution forauthentication; it combines the images with text and is a good solution for avoiding thebrute force attacks and to ensure a strong authentication scheme

3. Use the X.509 standards for obtaining the user credentials

Step 2 from the authentication scheme of Figure 6 suppose that when the user will registerinto the cloud service, he/she will receive a randomly grid of images. Each grid of images willcontain 3 images and each image will have a corresponding number (e.g. 1, 2, 3) (Figure 7).

The user will have to provide a secret code for each image. In this sense, the user will receive aregistration form, where it is asking to introduce secret characters for each corresponding numberlike in Figure 8.

Lets suppose that there were introduced the following codes like it is emplasized in Figure 8.After the user introduced their specific secret code for each corresponding image, the regis-

tration will be realized. The user should remember which code he/she had provided for each


Figure 6: Our CC SAP authentification Solution

Figure 7: Grid of images

Figure 8: The Registration Form


type of image (e.g. in the above example the user choose for the house image the hoho code,for the apple image the apap code and for the flower image the flfl code), because the proposedauthentication method requires entering these codes, but each time the images will be associatedwith different numbers (from 1 to 3), because the numbers are generated using a permutationalgorithm. Therefore, for authentication procedure there will be the same images each time, butwith another corresponding numbers (Figure 9).

Our hybrid text-image solution can be applied not only for accessing the cloud, but alsoas a authentication method at cloud client level (especially for mobile clients) and also at theapplication level.

Figure 9: Relationship between the randomly grid of images and the secret code provided byuser.

5 Conclusions and further work

A suitable authentication is required for organizations that want to access the Cloud services.Our solution regards increasing security at the Security Access Point level of Cloud Computingand it is in fact a strong hybrid user authentication solution based on using image combined withtext in order to avoid the weakness of simple user and password solution for authentication.

All authentication methods have drawbacks and currently there is not a system that cannotbe attacked. It is reasonable to assume that there is never likely to be a 100% secure system ofauthentication.

The biometric concept is extremely secure, but the biometric systems have the disadvantagethat require additional authentication periphery. This adds an additional cost that the standarduser is not willing to pay. Token-based authentication has seen a massive expansion in recentyears, especially in the banking sector. Adoption of smart card technology in the banking worldand access based on smartcards for access to companies and organizations have increased thedegree of usability of chip and PIN-based authentication.

Image-based authentication appears to offer the best solution. It provides increased security;is very versatile and does not require significant organizational changes in the enterprise. Interms of cost, image-based authentication is convenient as an alternative to text-based, because itrequire no significant extra costs. The main advantage of text-based authentication means that allapproaches of this kind of authentication are similar. Users are familiar with these authenticationsystems. In contrast, image-based approaches are likely to have different interfaces, which is likely


to be different from a security system of an organization to another one.A two factor password image based authentication method is proposed in this paper for

cloud services. This authentication approach is used without additional hardware involved andpresents the advantages of utilization in terms of security and usability. Every time when theuser will be asked to provide his/her identity, a form for each image included in the photo will belisted. The user will have to remember the secret code for each image and to carefully introducethem in the forms.

As a future work, we want to develop our approach for these randomizing appearing of imagesthat will increase the security level of the authetication system in cloud environment, being aneffective solution against shoulder surfing attacks.

Acknowledgement

This work was partially supported by the strategic grant POSDRU/88/1.5/S/50783, ProjectID50783 (2009), co-financed by the European Social Fund Investing in People, within the SectoralOperational Programme Human Resources Development 2007-2013.

Bibliography

[1] PARC,R.C., et al.,Authentication in the Clouds: A Framework and its Application to MobileUsers, ACM Cloud Computing Security Workshop (CCSW); 2010 October 8; Chicago, IL,2011.

[2] CSA, 2010. Domain 12: Guidance for Identity & Access Management V2.1. Cloud Secu-rity Alliance. Available at: http://www.cloudsecurityalliance.org/guidance/csaguide-dom12-v2.10.pdf, 2010.

[3] Metri P. and Sarote G., Privacy Issues and Challenges in Cloud computing, InternationalJournal of Advanced Engineering Sciences and Technologies, vol. 5, no. 1, pp. 5-6, 2011.

[4] Tianfield H., Cloud Computing Architectures, Proceedings of 2011 IEEE International Con-ference on Systems, Man and Cybernetics (SMC11), Anchorage, Alaska, USA, 2011

[5] Lonea A.M., Tianfield H., Popescu D.E., Identity management for cloud computing, In: NewConcepts and Applications in Soft Computing, Studies in Computational Intelligence Series,Volume 417, May 2012

[6] SETECS Inc, Security Architecture, for Cloud Computing Environments, White Pa-per, February 1, Available at: http://security.setecs.com/Documents/5 SETECS CloudSecurity Architecture.pdf, 2011

[7] Kay, R., Biometric authentication, retrieved April 20, 2005 2006 Availableat: http://www.computerworld.com/securitytopics/security/story/0,10801,100772,00.html,2006

[8] Tari, F., Ant Ozok, A., Holdon, H.S, A Comparison of Percieved and Real Shoulder-surfingRisks Between Alphanumeric and Graphical Passwords, retrieved June 10 2006 Available at:http://cups.cs.cmu.edu/soups/2006/proceedings/p56 tari.pdf, 2006,


[9] Dhamija R., et al, DĂŠjĂ Vu: a user study using images for authentication,Proceeding SSYM’00 Proceedings of the 9th conference on USENIX Security Sym-posium - Volume 9, USENIX Association Berkeley, CA, USA 2000, Available at:http://sparrow.ece.cmu.edu/ adrian/projects/usenix2000/usenix.pdf, 2000

[10] Jackson L., Analysis of Image-Based Authentication and its Role in Security Systems of theFuture, Available at: http://www.soc.napier.ac.uk/ bill/lee2006.pdf, 2006

[11] Nitin, Vivek Kumar Sehgal, et al., Image Based Authentication System with Sign-In Seal,Proceedings of the World Congress on Engineering and Computer Science 2008, WCECS2008, October 22 - 24, 2008, San Francisco, USA, 2008

[12] Newman R.E. HarshP., and Jayaraman P, Security Analysis of and Proposal for Image BasedAuthentication, IEEE Carnahan, 2005

[13] Confident Technologies Inc., Confident ImageShieldTM Available at:http://www.confidenttechnologies.com/products/confident-imageshield, 2011

[14] Renaud K., Just M., Pictures or Questions? Examining User Responses to Association-Based Authentication, to appear in the ACM Proceedings of the British HCI Conference2010, Dundee, Scotland, 6-10 September 2010.

[15] Just M. and Aspinall D., Personal choice and challenge questions: A security and usabilityassessment. In L. Cranor, editor, SOUPS, ACM International Conference Proceeding Series.ACM, 2009.

[16] Micallef N., Just M., Using Avatars for Improved Authentication with Challenge Questions,in Proceedings of the The Fifth International Conference on Emerging Security Information,Systems and Technologies (SECURWARE 2011), August 2011.


8(2):235-251, April, 2013.

Enhanced Daek Block Extraction Method PerformedAutomatically to Determine the Number of Clusters in

Unlabeled Datasets

P. Prabhu, K. Duraiswamy

Puinethaa PrabhuDepartment of Master of Computer ApplicationK.S. Rangasamy College of TechnologyTamil Nadu, India.Email: [email protected]

K. DuraiswamyDepartment of Computer Science and EngineeringK.S. Rangasamy College of TechnologyTamil Nadu, IndiaEmail: [email protected]

Abstract: One of the major issues in data cluster analysis is to decide the numberof clusters or groups from a set of unlabeled data. In addition, the presentation ofcluster should be analyzed to provide the accuracy of clustering objects. This paperpropose a new method called Enhanced-Dark Block Extraction (E-DBE), which auto-matically identifies the number of objects groups in unlabeled datasets. The proposedalgorithm relies on the available algorithm for visual assessment of cluster tendencyof a dataset, by using several common signal and image processing techniques. Themethod includes the following steps: 1.Generating an Enhanced Visual AssessmentTendency (E-VAT) image from a dissimilarity matrix which is the input for E-DBEalgorithm. 2. Processing image segmentation on E-VAT image to obtain a binaryimage then performs filter techniques. 3. Performing distance transformation to thefiltered binary image and projecting the pixels in the main diagonal alignment ofthe image to figure a projection signal. 4. Smoothing the outcrop signal, computingits first-order derivative and then detecting major peaks and valleys in the resultingsignal to acquire the number of clusters. E-DBE is a parameter-free algorithm toperform cluster analysis. Experiments of the method are presented on several UCI,synthetic and real world datasets.Keywords: Enhanced DBE, Automatic clustering, Cluster tendency, Visual assess-ment, Reordered dissimilarity image.

1 Introduction

The major concern in data mining is to outline the observed data into knowledge structures.Clustering aims at classifying objects of a related class into their relevant categories. Partitioningthe set of objects O = (o1, o2, ..., on) into C self-related objects is the major process of clusteranalysis. Various clustering algorithms are reported in the literature [1] and [2]. The generalproblems involved in clustering of unlabeled data sets are: a) assessing cluster tendency, i.e.,value of C. b) grouping the data into C meaningful sets and c) evaluating the discovered clustersC. This paper addresses the problem of determining whether the clusters are present by assessingof clustering tendency of clustering tendency as a prior process before clustering. Majority ofthe clustering algorithms need the number of clusters C as a key factor, so the quality of theresultant clusters mainly depends on the assessment of C.

Jain and Dubes [3] had discussed several statistically based informal techniques for clustertendency assessment. Ling [4] proposed a clustering algorithm based on estimated distribution


236 P. Prabhu, K. Duraiswamy

model. Cattell [5] formerly depicted pairwise dissimilarity information about a data set includingn objects as an n×n image, where the objects are suitably reordered so that the resultant image isimproved and is capable to emphasize the possible cluster structure in the data. The major papersin the visual representation of data dissimilarity include the contribution of [6], [7], [8] and [9].The universal denominator in all this methodology is reordered dissimilarity image (RDI). Theintensity of each pixel in the RDI represents the dissimilarity between the pair of objects denotedby the row and column of the pixel. An observer can merely calculate approximately the numberof clusters C (i.e., count the number of dark blocks along the diagonal) of an RDI where thedark blocks posse’s image lucidity (see Figure 1c).

Figure 1: An example for E-VAT image. (a) Scatter plot of a 3,000 - point’s data set with fiveclusters (b) Unordered image (c) Reordered E-VAT image I(D).

Generating RDIs could be done from any of the schemes anticipated in [6], [7], [8] [9] and [11].This paper develops a novel method to estimate automatically the number of dark blocks (seem-ingly also the number of possible clusters) in RDIs of unlabeled data sets. The proposed Enhanceddark block extraction (E-DBE) process combines several common images, signal processing tech-niques [10] and for the compactness, RDIs are generated using Enhanced Visual Assessment ofCluster Tendency (E-VAT) algorithm [11]. Later sequential image processing operations (region,segmentation, directional morphological filtering, and distance transformation) are performed tofragment the regions of interest in the RDI and then translate the filtered image into a distance-transformed image. Lastly, the altered image is projected on the diagonal axis of the RDI, whichyields an one-dimensional signal from which the (potential) number of clusters can be extractedfrom the dataset using signal processing operations.

The rest of this paper is structured as follows: In Section 2 we present the literature descrip-tion of visual approach. Section 3 reviews the enhanced VAT algorithm and Section 4 explainsthe procedure for Cluster Count Extraction (CCE) [12]. Section 5 analyses the dark block extrac-tion algorithm. Section 6 describes the proposed Enhanced DBE approach. Section 7 providesresults on UCI, synthetic and real world data sets for the proposed algorithm. The final sectioncontains a short discussion on results and for future study.

Enhanced Daek Block Extraction Method Performed Automatically to Determine the Numberof Clusters in Unlabeled Datasets 237

2 Literature review

Some existing approaches of the post clustering cluster validity problem are reviewed beforereciting the visual methods for cluster tendency assessment.

Index-based methods for cluster validity usually underline the intracluster density,intercluster division and additional factors such as geometric or statistical properties of the dataare proposed in [13], [14], [15], [16], [17], [18], [19], [20] and [24]. For instance, Milligan andCooper [13] compared 30 indices over a sequence of synthetic data sets. Above all these, Calinskiand Harabasz [14] index seems to be the best which performs the ratio between the traces ofthe between-cluster and within-cluster scatter matrix. It is a significant noting that the validityindices are completely dependent on the data and algorithm used to find partitions.

Probabilistic indices of cluster validity attempt to validate the number of clusters foundby probabilistic clustering algorithms. Guo [21] proposed a cluster number choice method fora small set of samples using a Bayesian Ying-Yang (BYY) model. Comparative studies suchas [17] and [22] provided experimental comparisons of many criteria such as Akaike’s InformationCriterion (AIC), Minimum Description Length (MDL), and (BYY) for determining the numberof clusters based on a Gaussian mixture model. A variety of statistical techniques for tendencyassessment are discussed in the work of Jain and Dubes [3].

Visual methods for cluster tendency assessment for a range of data analysis problemshave been extensively studied in [23]. Cattell [5] used single-linkage heuristics to rearrange theelements of small dissimilarity matrices, which were consequently hand-rendered for viewing.Floodgate and Hayes [7] offered hand-rendered pictures like Sneath’s, but reordering was donecomputationally using single-linkage clustering. Majority of the clustering algorithm builds RDIsprior to clustering and the RDI is viewed as a visual aid to tendency assessment. This is theproblem addressed by the new E-DBE algorithm, which uses the DBE algorithm of Liang [25]and E-VAT algorithm [11] to find RDIs and the number of clusters automatically.

A number of significant advantages of E-DBE over index-based or probabilistic methods aresummarized as follows:

• E-DBE is a preclustering technique, i.e., it does not need the data to be clustered, nordoes it locate clusters in the data. On the other hand, the consistency (and weakness)of postclustering index-based methods is entirely dependent on the clustering algorithmsused to identify the partitions.

• Index-based post clustering methods regularly need clustering to be performed severaltimes using a variety of cluster numbers and often find the top partition according to somepredefined criteria. Repetitive clustering can be computationally expensive, particularlywhen the range of possibe values of C remains uncertain. E-DBE has no such constraintand is performed just once

3 Review of Enhanced Visual Assessment Tendency

Of the many achievable ways to obtain an RDI, apply E-VAT to generate RDIs of unlabeleddata, i.e., to secure inputs to E-DBE algorithm. Let O = (o1, o2, o3...on) represent n objects in thedata. Vectorial data have the type F = (f1, f2, f3...fn), fi ⊂ Rh, where every coordinate of thevector fi provides an attribute value of each of h features (i.e., aj , j = 1, 2, 3...h) correspondingto an entity Oi. Constantly translate F into dissimilarities D = [dij = ||fi − fj||], 1 ≥ dij ≥ 0;dij = dji; dii = 0, for 1 ≤ i, j ≤ n. To make the paper self-sufficient, review of reorderingmethod E-VAT is shown in Table 1 which is proposed by [11] and an instance is shown in Figure1.


Table 1Enhanced-Visual Assessment Tendency Algorithm

InputConsider the dataset as n x n dissimilarity matrix.

D = [dij ]where1 ≥ dij ≥ 0; dij = dji; dii = 0, for1 ≤ i, j ≤ n

ProcessStep (1): Transform D to a new dissimilarity matrix R with dij = 1− exp(−dij/σ), where

σ is a scale parameter determined from D using the algorithm of Otsu [26] automatically.Step (2): Form an RDI image I(1) corresponding to R using the VAT algorithm [9].

Step (2.1): Let I = Φ, J = 1, 2, ...n and P = (0, .....0).Choose (i, j) ∈ argpjandq ∈j max {dpq}Place P (1) = i, I ← i and J ← J − {i}

Step (2.2): Iterate for t = 2...nSelect (i, j) ∈ argpiandq ∈j min {dpq}Set P (t) = j, revise I ← I ∪ {j} and J ← J − {j}

Step (2.3): Figure the dissimilarity template or matrix R = [dij ] = [dP (i)P (j)]Where 1 ≤ i, j ≤ n

Step (3): Display the reordered matrix R as the ODI I using the conventions given above.Output

Gray scale image I(D), which denotes maximum (dij) to white and minimum (dij) to black

Figure 1a shows the scatter plot of n = 3, 000 records points in R2, which are created froma combination of C = 5 bivariate normal distributions. These data points are transformedto a 3, 000 × 3, 000 dissimilarity matrix D by using distance measures for calculating distancebetween each pair of points. The five visually obvious clusters in Figure 1a are reflected by thefive separate dark blocks along the main diagonal in Figure 1c, which is the E-VAT image of therecords after reordering. On comparing with Figure 1b, which is the image of dissimilarities Din original input order, reordering is essential to expose the fundamental cluster structure of thedata.

The following are some points about E-VAT:

• E-VAT algorithm is performed to determine the number of clusters prior to clustering.Even if the estimated result does not match with the true value, it provides a basis forsetting the range.

• E-VAT depends merely on the input D, so a good quality D is decisive when D is a derivativeof object vectors. If the input dataset is of high dimensionality nonlinearly separable, itmay be improved by performing feature extraction.

4 Cluster Count Extraction (CCE) for Cluster Tendency Perfor-

mance

In the following sections, the performance of E-DBE is compared with other preclusteringassessment of cluster tendency techniques like DBE [25] and CCE algorithm [12]. CCE alsocounts dark blocks in RDIs using image transformation techniques. The major steps for thisalgorithm are summarized in Table 2.


Table 2The Cluster Count Extraction (CCE) Algorithm

Input

n×n - scaled matrix of dissimilarities D = [dij ] and its VAT image Image(D′) scaled so thatmax = white and min = black.

Step (1): Threshold Image(D′) with Otsu’s algorithm [26].

Step (2): Create a correlation filter ratio of size s′.

Step (3): Apply the Fast Fourier Transform (FFT) to both the segmented RDI and the filter.

Step (4): Proliferate tranformed VAT image with the composite conjugate of the transformedfilter.

Step (5): Compute inverse FFT for the filtered image.

Step (6): Acquire the off-diagonal pixel values (e.g., pth off-diagonal) of the back-transformedimage and calculate its histogram.

Step (7): Cut the histogram at an arbitrary horizontal line f = w and calculate the numeralof spikes.

Output

The number of dark blocks along the diagonal of Image(D′) called as C (Cluster Interger).

The CCE algorithm is applicable to built RDIs by any of the methods obtainable in the literature.In this algorithm VAT [9] was used to obtain RDIs from D, but E-VAT [11] is used in the proposedE-DBE algorithm. CCE algorithm is performed based on the parameter settings suggested in [12],i.e., s′ = 20, p = 1 and w = 0. Section 7 analyzes the results of CCE with DBE and proposedE-DBE on various synthetic, UCI Repository and Real-world datasets. The result in Table 5shows that E-DBE is more consistent than CCE because CCE algorithm performs on off-diagnalpixels values of the images which show the poor performance of the method.

5 Review of Dark Block Extraction (DBE)

Liang [25] proposed the Dark Block Extraction algorithm to estimate the cluster number inunlabeled data sets. DBE algorithm counts the dark blocks along the diagonal of an RDI usingbasic image processing techniques. The method is summarized in Table 3.

Table 3The Dark Block Extraction Algorithm

Input

n × n - scaled matrix of dissimilarities D = [dij ], the proportion of the allowed minimumcluster size of the data size n.

Step (1): Transform D to a new dissimilarity matrixD′ using σ - scale parameter determinedusing Otsu [26] automatically.

Step (2): Form an RDI image using VAT algorithm proposed by [9].


Step (3): Filter the image using morphological operators with directional line structuring ele-ments.

Step (4): Perform a distance transform on the image to obtain a new gray-scale image.

Step (5): Project the pixel values of the image onto the main diagonal axis of the image toform a projection signal.

Step (6): Smooth the obtained signal to get the filtered signal using filter techniques.

Step (7): Find peak positions Pi and valley positions Vj in the signal.

Step (8): Select major peaks and valley by removing minor ones using filters.

Output The numbers of dark blocks (i.e., the number of major peaks) are in the RDI.

The results of dark block extraction algorithm are displayed in Figure 2. Figure 2a shows thescatter plot of 3,000 points. These points are converted to a


pixel values of the DT Image(3) are projected onto the main diagonal axis to obtain a projectionsignal Histogram(1), as shown in Figure 3g. From the figure, C can be simply calculated becauseof the quite clear separation between major peaks in the signal Histogram(1).

Detection of major peaks and valleys in the projected signal (Steps 5-8): Theamount of dark blocks in any RDI is equal to the number of majorpeaks in the projection sig-nal Histogram(1). Based on the first − order derivative of the projection signal the clusternumber C is calculated from the detection of peaks and valleys. Although the projection signalHistogram(1) is available, need further smoothing to reduce possible false detections due to noisein the signal. Here Savitzky-Golay smoothing filters [29] (also called digital smoothing polyno-mial filters or least-squares smoothing filters) are typically used to smooth out a noisy signalwhose frequency span is large. In this algorithm, Savitzky-Golay smoothing filters perform muchbetter than typical averaging FIR filters performed in [25], which tend to filter out a significantportion of the signal’s high frequency content along with the noise. Savitzky-Golay filters areoptimal in the sense that they minimize the least-squares error in fitting a polynomial to framesof noisy data. It is well recognized that the peaks and valleys of a signal usually correspond tozero− crossing points in its first-order derivative, as shown in Figure 3g.

Remark - A significant issue for the E-DBE algorithm is how to successfully set the filter sizeα for the Savitzky-Golay filter. Actually, α is very simple to set because it reflects the minimumsupport threshold for the smallest cluster of importance in the data.

The novel in this algorithm is

• After preprocessing the dissimilarity matrix is transformed to a monotonic exponentialfunction.

• Distance measure used in this procedure is CityBlock distance which gives better results

• For better performance the proposed methods uses adaptive threshold for segmentation

• First order derivatives are computed

• For better projection of the signals the algorithm performs smooth, moving and savitzty-golay filters.

7 Experiment results of Synthetic, UCI and real world data sets

To assess the E-DBE algorithm with its prior measures, a number of experiments on severalsynthetically generated data sets, UCI Machine Learning Repository [30] as well as real-worlddata set are carried. The data sets’ characteristics and the results of CCE, DBE and enhancedDBE are accomplished in Table 5.

Table 5Summary of Synthetic, UCI and Real datasets’ distinctiveness and the results using CCE, DBE

and Enhanced-DBE


Data set #Instances

#Clusters#Eachcluster

Attributetype

#Attributes

CCE DBE E-DBE

SyntheticDatasets

Syntheticdataset -1

1000 2 [500,500] Integer 2 2 2 2

Syntheticdataset -2

1000 3 [500,250,250] Integer 2 2 2 3

Syntheticdataset -3

1800 3 [300,600,900] Integer 2 2 2 3

UCIDatasets

Dermatology 357 6 [110,59,70,48,51,19]

Integer 34 1 3 6

Heart 270 2 [150,120] Integer/Real

13 3 1 2

Hepatisis 72 2 [12,60] Integer/Real

20 1 1 2

Iris 150 3 [50,50,50] Integer/Real

5 1 2 3

Wine 178 3 [59,71,48] Integer/Real

13 2 2 3

Real worldDatasets

HIV 400 6 [221,144,11,17,5,1]

Integer/Real

19 1 3 6

7.1 Numerical examples with Synthetic Datasets

Observe the results on several synthetic datasets with multifaceted structures, in which anapparent cluster centroid for every cluster is not automatically available. Selections of syn-thetic datasets are based on the sets proposed in [25]. Synthetic Dataset (S − 1) is composedof two half-moon like patterns (C = 2). The dimension of the dataset is n = 1000, with 500points in each group. The upper half-moon is generated by fu(φ) = 2sin(φ) + 0.5 randn forφ = [π/500 : π/500 : π], while the lower half moon is produced by f1(φ) = 2sin(φ+ 0.6π) + 0.5rand for φ = [0.4π + π/500 : π/500 : 1.4π], where randn is a probability number drawn froma standard distribution with a zero mean and a standard deviation of one. Synthetic Dataset(S − 2) is generated from a grouping of two bivariate standard distributions and one half-moonlike model (C = 3). The magnitude of the data set is n = 1000, including 500 points for thehalf-moon pattern and 250 points for each of the two Gaussian shapes. The upper half-moon isgenerated by f(φ) = 2sin(φ) + 0.3 randn for φ = [π/500 : π/500 : π], where rand is a arbitrarynumber drawn from a regular distribution on the part interval. The two Gaussian shapes are gen-erated by the subsequent constituent parameters: the integration proporations are mean1 = 0.5and mean2 = 0.5; the mean values µ1 = (0.9, 0.5)T and µ2 = (2.1, 0.5)T ; and the covariancematrices

∑

1 =∑

2 = [10; 00.1].

Synthetic dataset(S − 3) is generated from a permutation of three circles with the identicalcentroid but diverse radii (C = 3). For every circle, generate synthetic data points by a1(φ) =


rsin(φ)+brandnsin(φ) and a2(φ) = rcos(φ)+brandncos(φ) where b is a constraint that controlsthe degree of overlap linking different circles, r is the radius of every circle, φ = [(2π)/p : (2π)/p :2π], and p is the size of each cluster. The synthetic datasets (S−1, S−2 and S−3) outcomes ofE-DBE algorithm using image processing techniques are depicted in Figure 4. The E-VAT imagesare shown in Figure 4a, Binary E-VAT images in 4b and the first order derivative ProjectionSignal obtained using smooth, moving and sgolay are presented in Figure 4c.

7.2 Numerical examples with UCI Machine Learning Repository

Next, consider some UCI datasets which are evaluated for the performance of proposed E-DBE method. The five datasets are dermatology, heart, hepatisis, iris and wine of UCI MachineLearning Repository [30]. For each dataset, the enhanced DBE with class attribute and dimen-sionality reduction [28] are performed. The UCI data sets’ characteristics and the consequencesof E-DBE are accomplished in Table 5.

Dermatology: The main intend of this database is to determine the category of Eryhemato-Squamous Disease. They all allocate the clinical features of erythema and scaling, with verymodest differences. The diseases in this group are psoriasis, seboreic dermatitis, lichen planus,pityriasis rosea, cronic dermatitis, and pityriasis rubra pilaris. The dataset include 357 occur-rences with 34 features including class attribute. i.e., 110 for class 1, 59 for class 2, 70 for class3, 48 for class 4, 51 for class 5 and 19 for class 6. Starting with 34-dimensional feature vectors,dataset are subjected to preprocessing, normalization and pairwise dissimilarities using the dis-tance measures to get relational data. Later the dissimilarity matrix D is submitted to E-DBEalgorithm for automatic clustering and the results are shown in Figure 5.

Heart: This dataset encloses the results of the prediction of heart attack. The dataset con-tains 72 instances and 13 attributes they are age, sex, chest pain type (4 values), resting bloodpressure, serum cholesterol and fasting blood sugar etc. The entire number of illustration in thisdata set is n=270, i.e., 150 represent absence and 120 the occurrence of heart attack. Initiallywith 13-dimensional feature vectors, the dataset preprocessing, normalization and pairwise dis-similarities by the distance measures are performed to acquire relational records. Afterwards, thedissimilarity matrix D is proposed to E-DBE algorithm for automatic grouping and the outcomeis depticted in Figure 6. Three clusters are shown as a result of CCE algorithm and one clusteris displayed as an outcome of DBE and two clusters by E-DBE (C=2).

Hepatisis: Hepatisis is an irritation of the liver characterized by the occurrence of inflam-matory cells in the tissue of the organs. This dataset contains the facts of the patient from


Figure 4: Results of the E-DBE algorithm on Synthetic datasets (S-1,S-2 and S-3) (a) E-VATImages synthetic data sets (b) Binary E-VAT images (c) First order derivative Projection Signalobtained using smooth,moving and sgolay.

Figure 5: Results of the E-DBE algorithm on Dermatology Dataset (a) E-VAT Image of Derma-tology Dataset, (b) Binary E-VAT image of Dermatology Dataset, (c) Distance Transformed Im-age, (d)First order derivative Projection Signal obtained using smooth, (e) First order derivativeProjection Signal obtained using moving, (f) First order derivative Projection Signal obtainedusing sgolay.


Figure 6: Results of the E-DBE algorithm on Heart Dataset (a) E-VAT Image of Heart Dataset,(b) Binary E-VAT image of Heart Dataset (c) Distance Transformed Image (d) First orderderivative Projection Signal obtained using smooth (e) First order derivative Projection Signalobtained using moving (f) First order derivative Projection Signal obtained using sgolay.

Figure 7: Results of the E-DBE algorithm on Hepatisis Dataset, (a) E-VAT Image of HepatisisDataset, (b) Binary E-VAT image of Hepatisis Dataset, (c) Distance Transformed Image, (d) Firstorder derivative Projection Signal obtained using smooth, (e) First order derivative ProjectionSignal obtained using moving, (f) First order derivative Projection Signal obtained using sgolay.


Figure 8: Results of the E-DBE algorithm on Iris Dataset, (a) E-VAT Image of Iris Dataset, (b)Binary E-VAT image of Iris Dataset, (c) Distance Transformed Image, (d) First order derivativeProjection Signal obtained using smooth, (e) First order derivative Projection Signal obtainedusing moving, (f) First order derivative Projection Signal obtained using sgolay.

respectively alcohol, malic acid, ash, magnesium, etc. The complete numeral of instances in thisitems are n =178, i.e., 59 for class 1, 71 for class 2 and 48 for class 3. The E-DBE results forwine data sets are shown in Figure 9.

Figure 9: Results of the E-DBE algorithm on Wine Dataset (a) E-VAT Image of Wine Dataset,(b) Binary E-VAT image of Wine Dataset (c) Distance Transformed Image (d) First order deriva-tive Projection Signal obtained using smooth (e) First order derivative Projection Signal obtainedusing moving (f) First order derivative Projection Signal obtained using sgolay.

7.3 Numerical example with Real-word Data set

The proposed method is tested on the HIV patient datasets collected from various Inte-grated counseling and Testing center (ICTC) and Antiretroviral (ART) centers of Tamilnaduand pondicherry. The preprocessing techniques are executed and then CCE, DBE and E-DBEalgorithms are applied to the HIV/AIDS diagnosis dataset containing 400 objects. Table 6shows the structure of the dataset with preprocessing depends upon the feature nature. Theattributes are respectively Age, Sex, WT, HB, Treat Drug, Pill count, Initial drug, Occupa-tion, Marital status, CD4, CD8, Ratio, WBC, RBC, PCV, platelet, TLC, SGPT, SGOP andDrug regimen- Class Attribute (CA). The complete numeral of items in this data set is n=400,i.e., 221 for class 1, 144 for class 2, 11 for class 3, 17 for class 4, 5 for class 5 and 1 for class


6. We computed pair wise dissimilarities using the Euclidean, Hamming, Mahalanobis dis-tance to get relational table. The E-DBE results shows the cluster count as five (C=6) whichis shown in Figure 10 a better result when compared with its prior algorithm CCE and DBE.Obj#

CA Age Sex HB WT Treat-Drug(regimen)

. . . CD4Count

WBC SGPT TLC

1 1 25 1 14 60 1 : 500 4600 46.0 4.02 2 35 1 11 48 2 : 100 6400 47.0 5.0: 1 : : : : : : : : : :: 1 : : : : : : : : : :400 2 45 0 13.5 58 1 . . . 150 3500 40.0 3.0

Figure 10: Results of the E-DBE algorithm on HIV- Drug Dataset (a) E-VAT Image of HIV-Drug Dataset, (b) Binary E-VAT image of HIV- Drug Dataset (c) Distance Transformed Image(d) First order derivative Projection Signal obtained using smooth (e) First order derivativeProjection Signal obtained using moving (f) First order derivative Projection Signal obtainedusing sgolay.

From the current study, the qualities of clusters are confirmed with the dark blocks on thediagonals and first order derivatives are achieved as peaks and valleys on the enhanced DBEcreation. It makes certain impact of objects related to the clusters in the reversed format.

8 Discussion and conclusion

This paper examines an almost parameter-free method for automatically estimating the num-ber of clusters in unlabeled data sets. The enhanced version of DBE algorithm works for un-specified data objects of n x n dissimilarity matrix and to estimate the feature of cluster beingdetermined. The only user-defined constraint that must be selected ? controls the filter sizefor applying filtering techniques. It is comparatively easy to make a pragmatic and functionalchoice for ?, since it effectively specifies the smallest cardinality of a cluster relative to the num-ber of objects in the data. The cluster number extracted by E-DBE appears to be increasinglyreliable. E-DBE will perhaps reach its useful limit when the RDI created by any reorderingof D is not from a well ordered dissimilarity matrix. In the proposed method distance metricsare explored for diverse types of given data sets which yield a better cluster visualization. Anachievable extension of this effort concerns the initialization of the c-means clustering algorithm


for object data clustering.Future work is proposed to obtain a visual clustering algorithm basedon the spectral analysis and E-VAT image and their distinctive block structured property to setthe data into C clusters. By mergeing cluster tendency assessment and cluster pattern usingan RDI, the proposed system can present a natural environment for visual cluster confirma-tion and analysis. To handle huge datasets, further propose a feasible approximate solution ina sampling plus extension manner to facilitate both visual cluster tendency estimation andpartitioning.

Bibliography

[1] R. Xu and D. Wunsch II, Survey of Clustering Algorithms, IEEE Trans. Neural Networks,16(3), 2005, 645-678.

[2] Shuliang Wang , Wenyan Gan, Deyi Li and Deren Li, Data Field for Hierarchical Clustering,International Journal of Data Warehousing and Mining, 7(4), 2011, 43-63.

[3] A.K. Jain, and R.C. Dubes, Algorithms for Clustering Data. Englewood Cliffs, (NJ: Prentice-Hall, 1988).

[4] Ling Tan, David Taniar, Kate A. Smith, A clustering algorithm based on an estimated dis-tribution model, International Journal of Business Intelligent and Data Mining, 1(2), 2005,229-245.

[5] R.B. Cattell, A Note on Correlation Clusters and Cluster Search Methods, Psychometrika,9(3), 1944, 169-184.

[6] P. Sneath, A Computer Approach to Numerical Taxonomy, J. General Microbiology, 17, 1957,201-226.

[7] G.D. Floodgate and P.R. Hayes, The Adansonian Taxonomy of Some Yellow PigmentedMarine Bacteria, J. General Microbiology, 30, 1963, 237-244.

[8] R.F. Ling, A Computer Generated Aid for Cluster Analysis, Comm. ACM, 16, 1973, 355-361.

[9] J.C. Bezdek and R. Hathaway, VAT: A Tool for Visual Assessment of (Cluster) Tendency,Proc. Int’l Joint Conf. Neural Networks (IJCNN ’02), 2002, 2225-2230.

[10] R.C. Gonzalez and R.E. Woods, Digital Image Processing (Prentice Hall, 2002).

[11] Puniethaa Prabhu and K.Duraiswamy, Enhanced VAT for Cluster Quality Assessment inUnlabeled Datasets, Journal of Circuits, Systems and Computers (JCSC), 21(1), 2012, 1-19.

[12] I. Sledge, J. Huband, and J.C. Bezdek, (Automatic) Cluster Count Extraction from UnlabeledDatasets, Joint Proc. Fourth Int’l Conf.Natural Computation (ICNC) and Fifth Int’l Conf.Fuzzy Systems and Knowledge Discovery (FSKD), 2008.

[13] G. Milligan and M. Cooper, An Examination of Procedures for Determining the Number ofClusters in a Data Set, Psychometrika, 50, 1985, 159-179.

[14] R.B. Calinski and J. Harabasz, A Dendrite Method for Cluster Analysis, Comm. in Statistics,3, 1974, 1-27.

[15] R. Tibshirani, G. Walther, and T. Hastie, Estimating the Number of Clusters in a Datasetvia the Gap Statistics, J. Royal Statistical Soc. B, 63, 2001, 411-423.


[16] U. Maulik and S. Bandyopadhyay, Performance Evaluation of Some Clustering Algorithmsand Validity Indices, IEEE Trans. Pattern Analysis and Machine Intelligence, 24(12), 2002,1650-1654.

[17] J.C. Bezdek, W. Li, Y. Attikiouzel, and M.P. Windham, A Geometric Approach to ClusterValidity for Normal Mixtures, Soft Computing, 1, 1997, 166-179.

[18] J.C. Bezdek and N.R. Pal, Some New Indices of Cluster Validity, IEEE Trans. System, Manand Cybernetics, 28(3), 1998, 301-315.

[19] W. Wang and Y. Zhang, On Fuzzy Cluster Validity Indices, Fuzzy Sets and Systems, 158,2007, 2095-2117.

[20] Decomposition Methodology for Knowledge Discovery and Data Mining, O. Maimon and L.Rokach, eds., 90-94, World Scientific, 2005.

[21] P. Guo, C. Chen, and M. Lyu, Cluster Number Selection for aSmall Set of Samples Usingthe Bayesian Ying-Yang Model, IEEE Trans. Neural Networks, 13(3), 2002, 757-763.

[22] X. Hu and L. Xu, A Comparative Study of Several Cluster Number Selection Criteria, Proc.Fourth Int’l Conf. Intelligent Data Eng. and Automated Learning (IDEAL ’03), 2003, 195-202.

[23] P.J. Rousseeuw, A Graphical Aid to the Interpretations and Validation of Cluster Analysis,J. Computational and Applied Math., 20, 1987, 53-65.

[24] Yun Sing Koh, Russel Pears and Gillian Dobbie, Automatic Item Weight Generation forPattern Mining and its Application, International Journal of Data Warehousing and Mining,7(3), 2011, 30-49.

[25] Liang Wang, Christopher Leckie, Kotagiri Ramamohanarao and James Bezdek, Automat-ically Determining the Number of Clusters in Unlabeled Data Sets, IEEE Transactions onknowledge and Data Engineering, 21(3), 2009, 335-350.

[26] N. Otsu, A Threshold Selection Method from Gray-level Histograms, IEEE Trans. Systems,Man, and Cybernetics, 9(1), 1979, 62-66.

[27] Mehmet Sezgin and Bulent Sankur, Survey over image thresholding techniques and quanti-tative performance Evaluation, Journal of Electronic Imaging, 13(1), 2004.

[28] Amit Saxena and John Wang, Dimensionality Reduction with Unsupervised Feature Selectionand Applying Non-Euclidean Norms for Classification Accuracy, International Journal ofData Warehousing and Mining, 6(2), 2010, 22-40.

[29] A. Savitzky and M.J.E Golay, Smoothing and differentiation of data by simplified leastsquares. Procedures, Analytical Chemistry, 36(8), 1964, 1627-1639.

[30] UCI Repository of Machine Learning Databases, http://www.ics.uci.edu/ mlearn /ML-Repository.html.


8(2):252-259, April, 2013.

Bio-Eco-Analysis for Risk Factors using GIS Software

R. Serbu, S. Borza, B. Marza

Razvan Serbu, Borza SorinBogdan MarzaLucian Blaga University of Sibiu10, Victoriei Bd., Sibiu, 550024, RomâniaE-mail: [email protected],[email protected],[email protected]

Abstract:Agriculture is a business sector ideally suited for the application of Geographic Infor-mation Systems (GIS) because it is natural resource based, requires the movement,distribution, and/or utilization of large quantities of products, goods, and services,and is increasingly required to record details of its business operations from the fieldto the marketplace. Nearly all agricultural data has some form of spatial component,and a GIS allows you to visualize information that might otherwise be difficult to in-terpret. Environment has a major impact on agriculture. In this paper we presentedhow GIS software can be used to analyze risk factors that influence agricultural pro-duction naturally. Natural risk factors were taken into account are: land degradation,flooding, humidity, action on farmland of the wildlife. The conclusions drawn fromthis paper using GIS allows the adoption of important measures on a short or longtime to reduce natural risk factors on agricultural production. The advantage of thismodel is possibility to bee extended to national, regional and global area.Keywords: Geographic Information Systems (G.I.S.), query, Information and Com-munication Technologies (ICT), spatial analysis.

1 Introduction

From several existing studies that debate relationship between environmental sustainability,economic performance and competitiveness has been debated strongly for many years and stillremains unclear. We have to bring to attention the Dr Javier Carrillo-Hermosilla’s book named"Eco-innovation". Here are the two main views that the literature [1], [2], [3], [4] gives us ofthe link between environmental and economic performance, which give rise to rather differentperspectives on this relationship. They are:

1. The ’traditionalist’, or neoclassical, view of a trade-off between environmental performanceand competitiveness. According to this view, the purpose of environmental regulation is to max-imize social welfare, making polluting firms responsible for the costs of the negative externalitythey produce, thereby correcting the market failure. As a consequence, environmental policiesmay have an adverse impact on competitiveness, insofar as this regulation imposes additionalcosts to firms. This burden may be of particular concern in industries with substantial environ-mental impact, where the share of environmental costs in total production costs is considerablyhigher than for the manufacturing sector on average. [5] A defensive business strategy and theadoption of end-of-pipe technologies may be expected. [6]

2. The ’revisionist’ view adopts a more dynamic perspective of the relationship between sus-tainability and competitiveness, and assigns a central role to technological change and innovation.Better environmental performance can lead to lower production costs and enhance competitive-ness through efficiency, productivity and new market opportunities. [7], [8], [9], [10], [11], [12]According to the so-called ’Porter Hypothesis’, [13] stringent environmental regulation could


Bio-Eco-Analysis for Risk Factors using GIS Software 253

force polluting firms to seek innovations to reduce the cost of compliance and production, im-proving the firm’s competitiveness and leading to a positive relationship between environmentaland economic performance. Additionally, companies can obtain ’first mover advantages’ by mar-keting the innovation itself and through the creation of new markets or market segments. [14]Hence properly designed environmental policies may help firms discover their inefficiencies andsources of comparative advantage, promoting innovation and creative thinking. [15] Setting asidethese theoretical discussions, and in more practical terms, it is clear that technological develop-ment and institutional considerations play an important role in the transition of the economicsystem towards sustainability. [16] In other words, technological change is probably a necessary,albeit insufficient, condition for achieving sustainability. Institutional changes, including changesin routines, social norms, formal regulations, etc., are needed not only to induce the requiredtechnological changes, but also to encourage behavioral changes at all levels of society in moresustainable directions.

Today’s major environmental problems, such as climate change, the destruction of the ozonelayer, loss of biodiversity, the degeneration and erosion of soil and water pollution are character-ized by their delocalization, considerable uncertainty, irreversibility and extreme complexity interms of consequences and the likelihood that they will occur. [17]

2 The analize of Environmental Risk in Agriculture with GISaplications

The analyze of Environmental Risk in Agriculture with GIS applications needs a multi-disciplinary approach, with input and expertise required from many fields - civil and chemicalengineering, physics, life sciences, ecology, geology, hydrology and statistics being some of them.A wide range of simple to complex, spatial as well as non-spatial, and quantitative as wellas qualitative, input data sets is used in environmental risk assessment and analysis process.The analyze of environmental risk in agriculture process involves preparation and use of theprocessed information derived and presented in various ways - for example, comparative (orrelative) risk analysis, cost-benefit analysis, scenario analysis, probabilistic analysis, decisionmatrix, sensitivity analysis etc. Due the need for using and analyzing a huge volume of thespatial as well as non-spatial environmental hazards and exposure data in a fast and reasonablyaccurate way, GIS based software applications using a variety of modeling techniques serve aspowerful tools for effective environmental risk assessment and management. [18]

Such applications can be used for a diverse environmental risk assessment and analysis pur-poses. These applications can ranges from development of databases/inventory systems for simpleto complex GIS layers overlays, to complex spatial decision-making systems for study of the im-pact of air, water and soil pollution, ecological imbalance, and natural disasters on the naturaland man-made environment, including living beings, properties, infrastructure, vegetation andecology. These systems could also be interlinked with other related systems, providing onlineand real-time input data feeds or communication systems, to allow continuous monitoring andtracking of environmental risks in an integrated way. Normally, it is good to start with a proto-type application first, which could be expanded further based on the budgetary allocation, userneeds and the user feedback obtained from the prototype’s implementation.

Such a system would allow the users to develop possible scenarios using GIS and graphicalicons For example, a symbol of a polluting industry planned can be placed at a user definedlocations on a given regional map, showing the terrain, rivers, soil, vegetation, population, em-ployment, infrastructure, land-use and wild life attributes, and to evaluate the different aspectsof the environmental risk, for a set of industry locations scenarios.

254 R. Serbu, S. Borza, B. Marza

It can help in developing a prior understanding of the potential risks and for arriving at thebest possible alternatives, within the given constraints. There actions could be a combination ofhuman actions taking place in that region; for example, establishment of an agricultural zone,clearing of some forest area to obtain news terrains for agriculture. This entire situation involvesa complex set of multiple actions over a wide geographic region, so there would a need for thesystem to be able to analyze & manage the environmental risks posed by the combination ofthese actions.

F. Capra states that by developing and using some systemic biology, each organic part of thecommon live is an integral whole and therefore a lively system regardless if we refer to individualsystems or social ones, to ecosystems we coexist with and we develop.

The economy as live organism is a system composed by human beings and social ones inter-acting with one another and the ecosystems that are around us and our life depend on.

It is from the perspective of the systemic vision that it comes out the understanding theproblems with interrelationships that are coming up at the level of the whole common liveeconomy is part of.

To look at the economic life apart from the environment, from people’s life, families andcommunities, from the life of the organizations and institutions means to fail understanding thateconomy is a live (vivid) system in a continuous change and evolution, dependent on the changeof the ecologic and social systems it finds it self related. It is from understanding the economy asa live (vivid) organism that a radical change results in the way the processes of economic growthand development are being conceived, mechanisms backed by the institutions that govern andmanage the crisis. This change resulted by the essence of systemic wisdom has the origin inunderstanding the wisdom of nature and is the substance of the ecologic consciousness Batesontalked about, the fact that our natural environment is not only live but conscious too.

A new economy as a live vivid organism entails a powerful investing in human resources, de-veloping the human wisdom in such a way as to enforce a new directing of science and technologytowards the organic, in a gentle, non violent, elegant way.

3 Spatial analysis of natural and hazards factors of risk in agri-culture using Geomedia Professional Software

The risk analysis of natural and hazard factors, we started with some definitions [19]:Hazard is "a threatening event or the likelihood in a region in a given period of a natural

phenomenon potentially harmful (damage, environmental damage, human casualties)Risk is defined as "the potential number of casualties, injuries, property damage of any

kind, produced during a reference period in a given region, where there is a particular naturalphenomenon".

Natural disaster is "a serious disruption of functioning of a society, causing loss of life, ma-terials and environment, which the company can not exceed in-house".

The vulnerability is "the degree of loss (0-100%), a phenomenon resulting from potentialityto produce casualties and material damage", depending on the vulnerability of socio-economicdevelopment of the area concerned.

In addition to the above definitions, there are many approaches that are intended to comple-ment and enhance the significance of the terms set [20].

Analysis of risk factors in Sibiu County was based on a map using GIS software, GeomediaProfessional. We obtain the map with major risk factors [21].

The Areas where they occur are presented in map:


Figure 1

• Relatively stable areas with different risk of flooding and embankment works without clog-ging and regulate rivers and streams of water and maintenance of beds;• Stable areas but with the risk of water stagnation, due to low permeability of soil;• Risk areas due to excessive soil moisture by raising the groundwater or irrigation;• Moderate to strong erosion areas with high risk of landslides activation when heavy rain,

deforestation or work on slopes;• Strong unstable areas affected by erosion, excessive, coupled with active gulling and land-

slides, torrents, and springs coastal;• Unstable areas at high risk of landslides, collapses and collapsing;• Seismic zone, (MSK scale);• Area of seismic risk as normative P100-1992 (F=0,008;E=0,12;D=0,16)• Areas affected by landslides;• Areas affected by floods due to overflowing rivers;• Areas affected by floods due to leakage from the slopes.The legend shown in the figure above, appears in Figure 2.All these risk areas are represented in the map made. GIS product provides great opportu-

nities for spatial analysis [22]. You can see the weight of each risk factor in the whole county.

The analysis of the digital map shows that:• High risk factor for landslides is relatively high in Sibiu county, this having a negative

influence on agriculture and human habitat;• Soil erosion with negative influences on agriculture is another important factor which covers

a large area in the county of Sibiu;• Flood risk is important in the county. On the map are located in areas where floods

occurred with disastrous effects. In Sibiu floods occurred in the years: 1956, 1958, 1960, 1962,1965, 1967, 1969, 1970, 1971, 1975, 1982, 1985, 1998, 2005;• In general the county is a hilly area, the beds of watercourses characterized by gently sloping

broad flood plain areas, without equipped with storage or flood defense and without extraction


Figure 2

opportunity ballast;• The analysis of the map also notes the significant share that has seismic risk factor in Sibiu

County.GIS software allows obtaining custom information using spatial analysis and the particular

attribute and spatial queries [23]. So you can analyze the influence of risk on a certain area,using buffer zones. For example in the figure below we present a buffer area for flood risk factor.

In the map made, we may introduce new risk factor such as, referring to the influence ofwildlife on agricultural production. Although the number of wild animals in Sibiu County de-creased by about 50% in last years, it continues to represent a significant risk on agriculturalproduction, in rural areas

Given the issues outlined above, Figure 4 presents a map that allows the analysis of othernatural factors of risk. The map has 3D features.

4 Conclusions and Future Works

Spatial analysis is very importance to all areas and in agriculture also. The spatial analysiscan determine the relevant institutions to take certain measures to be taken as:• Make works of dams and abundant vegetation cleaning;• Reforestation areas with landslides;• Cant dams;• Repair of retaining walls;• Desalting of water courses;• Desalting of culverts sections.Motivations of people who live within rural areas, to adopt and use information technologies

and those of communication, can differ from the reasons people have within the urban environ-ment. Regarding the implications of Geographic Information Systems in bio-economic analyses


Figure 3

Figure 4


of some natural factor of risk, the people from rural area should be more interested because nowthey can minimize the gap between rural and urban development.

Researches suggest that information and communication technologies (ICT) can eliminatethe handicap of distance concerning distance and social interaction. Donald Janelle used theexpression "convergence time-space at the end of year ‘60 in order to describe the capacity oftransport, of technologies, in order to approach different places" [18]. The risk we discus about,are between the one that minimize the profit and the good living of people living in the ruralarea. These people can take advantage of these technologies, and information and communicationtechnologies (ICT) offer this way entirely.

Using information and communication technologies might have greater impact on some per-sons within the rural area, than another one who lives in the urban environment, because aperson who lives in rural area can now access information, goods and services and very impor-tant, information that they couldn’t before.

Now when the global crisis is being considered as a complex, multi dimensional one with facesthat reach each aspect of our life - health and living conditions, quality of the environment andsocial, economic, technological, political relationships... a crisis of intellectual, moral, spiritualdimensions , a vast crisis as there has not been before in humanity, human re- spirituality assustainable ground of that of the institutions representing a redefining of the human nature fromthe perspective of the fact the revolution of the means must be subordinated permanently toexpectations , only out of their harmony results the health of the entire common live formed ofpeople, communities, organizations, families and institutions.

Acknowledgements

This paper was co-financed from the European Social Fund through Sectoral Operational Pro-gramme Human Resources Development 2007 - 2013, project number POSDRU /89/1.5/S/63258"Postdoctoral school for zootechnical biodiversity and food biotehnology based on the eco-economyand the bio-economy required by eco-san-genesys."

Bibliography

[1] Xepapadeas, A. and De Zeeuw, A. Environmental policy and competitiveness: The Porterhypothesis and the composition of capital, Journal of Environmental Economics and Man-agement, 37, 165-182, (1999);

[2] Simpson, R. D. and Bradford, R. L. Taxing variable cost: Environmental regulation asindustry policy, Journal of Environmental Economics and Management, 30, 282-300, (1996);

[3] Palmer, K. W., Oates, W. E. and Portney, P. R. Tightening environmental standards: Thebenefit-cost or the no-cost paradigm, Journal of Economic Perspectives, 9(4), 119-132, (1995);

[4] Walley, N. and Whitehead, B. It’s Not Easy Being Green, Harvard Business Review, 72(3),36-44, (1994).

[5] Luken, R. The Effect of Environmental Regulations on Industrial Competitiveness of SelectedIndustries in Developing Countries, Greener Management International, 19, 67-78, (1997).

[6] Faucheux, S. and Nicolai, I.,Les firmes face au development soutenable: changement tech-nologique et gouvernance au sein de la dynamique industrielle, Revue d’Economie Indus-trielle, 83, 127-145, (1998).


[7] Sinclair-Desgagné, B. Remarks on Environmental Regulation, Firm Behaviour and Innova-tion, Scientific Series 99s-20, 1999, (Montreal: Cirano);

[8] Porter, M. and Van der Linde, C. Green and Competitive: Ending the Stalemate, HarvardBusiness Review, September/October 1995, 120-134, (1995a);

[9] Porter, M. and Van der Linde, C. Toward a New Conception of the Environment-Competitiveness Relationship, Journal of Economic Perspectives, 9.4, 97-118, (1995b);

[10] Shrivastava, P. Ecocentric Management for a Risk Society, Academy of Management Review,20.1, 118-37, (1995);

[11] Porter, M. America’s Green Strategy, Scientific American, 264.4, 96, (1991);

[12] Gabel, L. H. and Sinclair-Desgagné B. Managerial Incentives and Environmental Compli-ance, Journal of Environmental Economics and Management, 24, 940-55, (1993).

[13] Porter, M. and Van der Linde, C. Green and Competitive: Ending the St, alemate, HarvardBusiness Review, September/October 1995, 120-134,(1995a).

[14] Esty, D. and Porter, M. Industrial Ecology and Competitiveness: Strategic Implications forthe Firm, Journal of Industrial Ecology, 2.1, 35-43, (1998); Shrivastava (1995) op. cit.

[15] Jaffe, B., Peterson R., Portney R. and Stavins R. Environmental Regulation and the Com-petitiveness of US Manufacturing: What Does the Evidence Tell Us?, Journal of EconomicLiterature, 33, 132 -63, (1995).

[16] WCED, Our Common Future (Oxford University Press for the World Commission on En-vironment and Development), (1987).

[17] Dr Javier Carrillo-Hermosilla’s, Dr Pablo del Río González, Dr Totti Könnölä, Eco-innovationt, Palgrave Macmillan, 2009.

[18] Pierce , F., Clay, F., GIS Applications in Agriculture, CRC Press, Taylor & Francis Group,2007.

[19] Bryant E.A., Natural hazards, Cambridge University Press, 1992.

[20] Tanislav D., Costache A., Geografia hazardelor naturale şi antropice, Editura Transversal,Târgovişte, 2007.

[21] Borza S., Realizarea Aplicaţiilor GIS, folosind Geomedia Professional, Ed. Universităţii "Lu-cian Blaga" din Sibiu, ISBN 978-606-12-0261-4, 2012.

[22] Bălteanu D., Natural hazards in Romania, "Revue Roumaine de Géographie", 36, Bucureşti,p. 47-55, 1992.

[23] Donald G. Janellea, Spatial reorganization: a model and concept, Annals of the Associationof American Geographers Volume 59, Issue 2, pages 348-364, June 1969.


8(2):260-273, April, 2013.

Evaluation of an Information Assistance System Based on anAgent-Based Architecture in Transportation Domain: First

Results

A. Trabelsi, H. Ezzedine

Abdelwaheb Trabelsi1. LOGIQ-GIADE, FSEG, University of Sfax2. Faculty of sciences, PB 1171, 3000 Sfax, TunisiaE-mail: [email protected]

Houcine Ezzeddine1. Univ Lille Nord de France, F-59000 Lille, France2. UVHC, LAMIH, F-59313 Valenciennes, France3. CNRS, UMR 8530, F-59313 Valenciennes, FranceE-mail: [email protected]

Abstract:The evaluation of interactive systems is a wide and rich research domain: many meth-ods, criteria and tools are available. In this article, we are focused on agent-basedinteractive systems. We first describe the agent oriented architecture used in ourresearches. Then we propose an evaluation approach based on three complemen-tary techniques: assistance evaluation tool, questionnaire, and verbalization. Thevalidation of our approach occurred within the framework of a project involving anindustrial partner which is running the current urban transport network (tramwayand bus) in the town of Valenciennes, France. The main results of the evaluation ofan agent-based Information Assistance System (IAS) are presented. This evaluationis based on two scenarios: normal or distrupted running mode. This evaluation hasbeen conducted in laboratory.Keywords: Evaluation, Human-Computer Interaction, Electronic informer, Infor-mation Assistance System, traffic regulation.

1 Introduction

The evaluation of interactive system consists in ensuring that the users are able to carryout their task by using the system; it must therefore meet their needs. The methods and toolscurrently available for evaluating interactive systems are numerous and various; we quote forinstance: observations, eye tracking, interviews, electronic informers, questionnaires, user tests,inspection methods, knowledge based automated systems, and so on [14], [19], [16], [20]). Eachof them presents some advantages and drawbacks.In this article we are particularly interested in the evaluation of agent-based interactive systems.Indeed agent-based architectures of interactive systems have been proposed since the eighties inthe literature. Such architectures lead to new needs concerning the evaluation of the interactivesystems concerned [15], [28], [5], [8].Thus, we propose an evaluation approach based on three complementary techniques: assistanceevaluation tool, questionnaire and verbalization. Urban transport networks, and in particular thesystem which provides information for the passengers (IAS: Information Assistance System [6]),will be used as an example of the application of our approach.In this article, we will present the main results obtained from the evaluation of the Informa-tion Assistance System (IAS). Indeed, the evaluation in the laboratory has initially enabled usto technically test the proposed assistance evaluation system, and secondly to detect a priorrepresentative set of utility and/or usability problems, inherent to the IAS exploitation.


Evaluation of an Information Assistance System Based on an Agent-Based Architecture inTransportation Domain: First Results 261

2 Agent oriented architecture for interactive systems

The architectural design of interactive systems is object of many researches since the eighties.Architecture of an interactive system is composed by components, the outside visible propertiesof these components and the relations between them [4]. In general, the proposed models respectthe following principle: the separation of the user interface part from functional core (application)part; as a result, the flexibility, reusability and maintainability are increased. We can distinguishtwo types of architectural models:

• Functional models, such as the Seeheim and Arch models [1]; functional models split aninteractive system into several functional components; for instance, the Seeheim model ismade up of three components (Presentation, Dialogue Controller, Application Interface).

• Structural models, such as PAC [3], AMF [2] or MVC [11] (and their variations); the struc-tural models aim at a finer breakdown. Indeed, such models regroup functions together intoone autonomous and cooperative entity (often called agent). They are agent-based inter-active systems that are built based on a hierarchical structure of agents in accordance withthe principle of composition or communication (not on a functional division like functionalmodels). For example, PAC model is a hierarchical structure of interactive agents: a PACagent is composed by three facets: Presentation that connects agents to the input/outputdevices, Abstraction is responsible of functional core of the application, Control plays anintermediary role between the two other components and serves communications betweenPAC agents. Three facets (Model, View and the Controller) also compose an agent of theMVC model.

Each of these models has its own advantages and disadvantages. In order to exploit the ad-vantages of both types of models, we propose an agent-based architectural model that borrowsprinciples of both of them; so this model can be considered as a mixed model. The idea of amixed model is not new but our proposed agent-based architectural model aims principally at(1) designing complex supervision systems in industrial context, (2) proposing solutions for theevaluation phase, as explained in [25] and [30].In the architecture used in our researches, we suggest using a division into three functionalcomponents (see Figure. 1) which we have called respectively: interface with the application(connected to the application), dialogue controller, and presentation (this component is directlylinked to the user). These three components group together agents:

Figure 1: Proposed agent-based architectural model of interactive systems [8]

• The application agents handle the field concepts; they cannot be directly accessed by theuser. One of their roles is to ensure the correct functioning of the application and the realtime dispatch of the information necessary for the other agents to perform their task,

• The dialogue control agents are also called mixed agents; these provide services for both theapplication and the user. They are intended to guarantee coherency in the data exchangesemanating from the application towards the user, and vice versa,

262 A. Trabelsi, H. Ezzedine

• The interactive agents (also called interface agents), unlike the application agents, are indirect contact with the user (they can be seen by the user). These agents coordinate be-tween themselves in order to intercept the user commands and to form a presentation whichallows the user to gain an overall understanding of the current state of the application. Inthis way, a window may be considered as being an interactive agent in its own right; itsspecification describes its presentation and the services it has to perform.

3 Agent oriented specification and design of the information as-sistance system

Agent oriented architecture has been used for the design of the first version of an InformationAssistance System (IAS) [8], [9]. The application agents are intended to manage the passengerinformation in the vehicles and stations and to calculate the information to be displayed (delays,timetable and route modifications, etc.). Thus we can consider the IAS as a complex systemwhich is a very rich research and development field [10]. According to the traffic context, eachagent possesses rules enabling it to act correctly in its environment. Concerning the specificationof the interface agents, we have identified six types of interface agent responsible for directinteraction with the user (human regulator). They are represented in the form of interactivewindows. The user can interact with them via the various functions possible in the windows:buttons, edition zones, pictures, and so on. These agents are:

• The State of the traffic interface agent: it gives a synthetic representation of all the delaysconcerning mobile units travelling on the network. Thus, with the help of the networksupport system, it ensures the real time surveillance of vehicle delays on the networksupervised.

• The State of the line interface agent: the view of this agent is made up of graphic elementssuch as stations, route sections, vehicles, and so on (see Figure. 2a). A click on a vehicledirectly displays the view (window) of the Vehicle agent which will deal with any furtherinteraction with the user (see Figure. 2b). The principle is the same when the user clickson a station (see Figure. 2c).

• The Station and Vehicle interface agents: the view of these two agents is accessible byacting on their associated representations in the State of the line interface agent view(vehicle and station). It shows the user the information contained in the running plansin the form of a set of thumbnails depending on a direction which can be selected on adrop-down scroll list.

• The Message interface agent: it enables the human regulator to obtain a synthetic view ofall the messages being sent to vehicles and stations.

• The Overall View interface agent: in order to make the task of supervising the traffic easierfor the regulator, we created it. As its name implies, the view given by this agent providesthe user with a global view of the traffic on the network. This view encompasses all thelines to be supervised and facilitates access to line, stations and vehicles.

A survey of the user interfaces related to these agents is available in [23].


Figure 2: (a) View of the State of the line interface agent, (b) the view of a Vehicle interfaceagent,(c) View of a Station interface agent [27], [23]

4 Evaluation of the information assistance system (IAS)

4.1 Experimental device

The experimental device of the IAS evaluation consists of three techniques and tools: anevaluation assistance system integrating an electronic informer called MESIA1 [26], questionnaireand verbalization. The use of each of them is presented hereafter.

Evaluation assistance system

The evaluation assistance system is composed of several modules, shown in the middle ofFigure 3 [24], [27].The electronic informer module is directly connected to the interactive system to be evaluatedby the association of an informer agent to each agent of the interface. The creation of theseinformer agents is deduced directly from the architecture of the system which is to be evaluated,more specifically from the presentation agent-based system.Once the interaction data has been collected and stored, it is used by a module able to generate atask model. This is based on the exploitation of agent Petri nets, inspired by parametrized Petrinets [12], selected for their ability to handle entities of the agent type, according to principlesdescribed in [9]: the model obtained corresponds to that of the real activity. This module isalso able to generate a model corresponding to the task to be performed, whose components areavailable in a base intended for this purpose (stored in the BMT(R) base, cf. below). Indeed,two bases are available [22], [27]

• The Base of Specifications of Agents (BSA) allows the storage of the specifications of theinterface agents. It contains the definition (for each agent) of the sets E (set of the possibleevents), C (set of the conditions), R (set of the resources), Acv (set of the visible actions:such as the action of the user using the mouse or the keyboard, the reaction of the interfaceby the posting of new windows and/or change of their contents), Acn (set of the actionswhich are not visible to the user, relating to the interactions between interface agents).

1Mouchard électronique dédié à l’Evaluation des Systčmes Interactifs orientés Agents; translated by: Electronicinformer dedicated to the Evaluation of Agent-Based Interactive Systems.


The data stored in BSA is intended to be exploited by the module of task model generation(taking the form of agent Petri nets).

• The Base of Task Models (BMT) is composed of two sub-bases called BMT(O) andBMT(R). BMT(O) contains the description of the task observed, the models being gener-ated by the module of generation (of Petri nets). BMT(R) contains the description of thetasks, also called prescribed or reference tasks (to be realized by the users), such as theyare described a priori by the designers or evaluators via a module allowing Simulation/-Confrontation/Specification of agent Petri nets (cf. below).

The Simulation/Confrontation/Specification module provides the evaluators/designers with thefollowing three functionalities:

• Simulation of agent Petri nets: this function allows the visualization of agent Petri netsdynamics, and in consequence provides an overview concerning the HCI dynamics; this isbecause of the exploitation of the task models (modelled by agent Petri nets) and of theformulation which ensures the evolution in agent Petri nets.

• Confrontation of agent Petri nets: this function exploits the task models (Observed, ofReference) for confrontation (according to the principles described in [7]). This confronta-tion aims make it easier for the evaluators to identify possible ergonomic problems relatedto the usability of the interactive system; for example to realize that agent Petri nets ofthe task model observed contains states in which, for example, the user passes by uselessstages, or where the time taken to carry out a task is far greater than that envisaged apriori by the evaluators/designers.

• Specification of agent Petri nets: this function consists of providing the evaluators/designerswith means (windows) allowing the management (description, modification, ...) of the agentspecification, in other words the definition of the E, C, R, Acv, Acn sets and their storagein the agent specification base (BSA).

More explanation about a specific tool ensuring the generation, the simulation and the con-frontation of Petri nets can be found in [30], [29], [28].In addition to the verbalization, the use of the electronic informer as a tool for the evaluation ofthe IAS is enriched by subjective answers to a questionnaire.

Questionnaire

The questionnaire prepared specifically for this evaluation is basically inspired from [21]. It iscomposed of three parts: the first presents questions about the user interface general aspects (suchas the response time); the second gathers specific questions about each IAS view (window); andthe third part presents a global ergonomic evaluation of the user interface. The complementarytechnique to the questionnaires is the verbalization.

Verbalization

The verbalization is an easy and direct means to collect information about the quality of thesystem and particularly the user interface. Contrary to the questionnaire, the verbalization hasthe advantage of being more flexible insofar as it allows the orientation of the questions towardsthe information sought by the evaluator [13].


Figure 3: Experimental device used for the IAS evaluation

4.2 Population implied in the evaluation

According to [17], [31], it is possible, with 4 or 5 participants, to detect nearly 80 to 85% ofthe utilisability problems. Hence, for this first evaluation of the IAS in laboratory, the populationis made up of five subjects. They have an average age of 29 years; all are male. All the subjectsare PhD students in computer science. We have considered that the fact that the subjectsare experts in computer science can compensate, to some extent, their lack of experience inregulation. Indeed, the familiarity of PhD students with the computer software and the human-machine interfaces handling can put them draw sheet with the use of the IAS.However, it is clear that the results obtained within the framework of this first evaluation willnever be as consistent as those which could be obtained with professional human operators inregulation rooms.

4.3 Experimental protocol

Each experiment lasts approximately an hour and half and comprises four phases which arepresented in Table 1. During the first phase of the experiment, which lasts fifteen minutes, theIAS and its global functioning are explained to the subject. The subject is then familiarizedwith the various views of the IAS during approximately ten minutes before undergoing truly theexperimental tests. The right handling proceeds in twenty minutes and relates to two scenarios.Ten minutes are devoted to each scenario. The last phase of the evaluation relates to the fillingof the questionnaire and the verbalization; it lasts approximately forty-five minutes.


Table 1: Phases of the evaluation

Phase duration Task description

1 15 minutesReception, description of the evaluation objectives, and global pre-sentation of the Information Assistance System (IAS)

2 10 minutes Learning and free trial of the IAS3 20 minutes Realization (by the user) of two previously prepared scenarios4 45 minutes Responses to the questionnaire and verbalization of the user

4.4 Tasks to be realized

Scenario 1: evaluation of the IAS in normal running mode

In the reality, in normal running mode, analyses showed that the task of the human regulatorsamounts to the supervision of the traffic, but that they can also, of their own initiative, sendmessages to vehicles and to stations. Within the framework of the evaluation, we thus get closer tothis established fact. To make sure that the user (the subject) easily succeeds in interacting withthe IAS by sending messages to the station(s) and the vehicle(s), we propose a first evaluationscenario made up of four tasks described in Table 2.

Table 2: First scenario: tasks to be realised

Tasks to be Theoretical durationTask description

realized of the task

T1 45 secondsSend a message to the station " Gare SNCF "of the tramway line: Stop of the next tramway2 minutes in the station

T2 45 secondsSend a message to the tramway driver N◦ 6:Stop 2 minutes in the next station

T3 60 secondsSend a message to all the stations of line 16:Disrupted traffic because of a manifestation

T4 60 secondsSend a message to all the vehicles: Merry hol-idays

Let us note that theoretical durations necessary to the realization of the tasks are approximatedurations. They are determined by a supervision expert initiated with the SAI.

Scenario 2: evaluation of the IAS in disrupted running mode

In the reality, in disrupted running mode, analyses showed that the task of the humanregulators amounts to react by regulation actions to the warning, abnormality or breakdownmessages from the system. By regulation action, we mean the information sent by the regulatorand which is intended for the vehicles drivers and for the travelers in both stations and vehicles.

To make sure that the user (subject) interacts easily with the IAS by making regulationactions, we propose the second scenario which is complementary to the first one, and composedby four tasks described in Tables 3 and 4.

It is noticeable that the duration between the appearances of two messages is relatively short;this is not a fate. Indeed, we wish to put the user in a situation close to the reality where several


Table 3: Second scenario: messages displayed to the subject

Time Kind of messageMessage from the Exploitation AssistanceSystem (EAS)

t = 15 seconds warningMessage 1: the vehicle N◦ 4 line 16 is in advanceof 5 min

t = 35 seconds warningMessage 2: the vehicle N◦ 2 line 17 is late of 5min

t = 1 minute breakdownMessage 3: the vehicle N◦ 4 line 16 is out oforder

t = 2 minutes abnormalityMessage 4: Incident on the line Tramway nearthe Gare-SNCF station

Table 4: Second scenario: tasks to be realized

Tasks to be Theoretical durationTask description


T1 55 secondsSend a message to the vehicle driver N◦4 line 16:Stop 2 minutes in the next station

T2 55 secondsSend a message to the vehicle driver N◦2 line 17:You are late, please accelerate if possible

T3 2 minutes

• Send a message to the passengers of vehicleN◦4 line 16: Out of order bus, arrival of thenext bus in 15 minutes

• Send a message to the vehicle driver N◦4line 16: The breakdown service arrives in10 minutes

• Remove the vehicle from the network

T4 1 minuteSend a message to concerned stations: Disruptedtraffic: accident on the tramway line

incidents can arise simultaneously within the network.Contrary to the first scenario, the execution of the tasks is not sequential. Indeed, the beginningof each task is announced by an alarm or warning message.In every appearance of a message of the Table 3, the user has to perform the task of the corre-sponding regulation. The Table 4 shows the tasks to be realized for each received message. Thetheoretical durations necessary for the realization of the tasks are determined by an expert insupervision initiated to the IAS.

5 Results

The five subjects carry out the various tasks envisaged with the two scenarios. The experi-mental device used enables us to collect (1) objective data by means of the electronic informer(MESIA) and (2) subjective data via the verbalization and the fillings of the questionnaire. We


present in what follows some relevant results.

5.1 Results of the first scenario

The Table 5 presents a summary of the real average duration of each task in the first scenarioas well as the success rate of its realization.

Table 5: Result of the first scenario

Tasks to be Real average durationSuccess rate of the task realization


T1 41 seconds 100 %T2 39 seconds 100 %T3 67 seconds (3 subjects out of 5)T4 75 seconds (3 subjects out of 5)

The results displayed in this Table 5 indicate that all the subjects carried out well the firstand the second task with an acceptable realization average time; whereas, for the third andfourth task, only three subjects out of five could complete them. We also notice that the realrealization duration of these two tasks exceeds the average. Indeed, this can be explained by thefact that the subjects are not experts in regulation.To understand better the results obtained in Table 5, we can compare the model of the performedtask and the model of the task to be realized. We take as an example the task T3 presented inTable 4.Figure 4 shows:

• the model of the task to be realized (part a),

• the model of the performed task, successfully, by the subjects 1,3 and 5 (part b),

• the performed task, with failure, by the subject 4 (part c),

• the performed task, with failure, by the subject 2 (part d).

The Petri nets presented in Figure 4 show that the subjects 2 and 4 failed to achieve their task.Indeed, both subjects cannot reach the view Message. The subject 2 is blocked in the viewStation and the subject 4 is blocked in the Vehicle view. This report confirms the results ofTable 5.This problem of blocking can be seen as a usability problem. Indeed, the results obtained afterthe evaluation of the IAS with the first scenario reveal that the IAS does not allow an intuitiveand easy access to the view Message. We shall see farther some improvements related to theState of the line view which aims to introduce a specific message sending zone (cf. figure 5).This zone will allow the user to reach directly the Message view without having to access to theStation view or the Vehicle view.

5.2 Results of the second scenario

The five subjects realized the four tasks foreseen in the second scenario. Table 6 (obtainedthanks to the electronic informer) presents a summary of the real average duration of every taskas well as the rate of success of its realization.


Figure 4: The model of the task to be realized and of the performed task (task 3, scenario 1)

Table 6: Result of the second scenario

Tasks to be Real average durationSuccess rate of the task realization


T1 1 minutes 10 seconds 100 %T2 1 minutes 30 seconds 100 %T3 3 minutes 20 seconds 100 %T4 1 minutes 25 seconds 100 %

The obtained results show that the five subjects succeed to achieve the four tasks proposed.However, we note an overtaking of theoretical time foreseen to perform the four tasks to be real-ized. This observation finds an explanation in the collected data with the verbalization. Indeed,all the subjects, without exception, point out that it is impossible to them to memorize messagesfrom the exploitation assistance system (system in which the position and state of each vehicleare stored [6]). Besides, the IAS jams until the user validates the message; the user is thusobliged to memorize the message or to note it.Besides, the Petri nets reconstruction of the performed tasks did not reveal any particular prob-lem.Other results are available in [23]. They lead to several improvements resumed below.

6 First IAS improvements

Regarding the results of the IAS evaluation obtained with the two suggested scenarios, theanswers to the questionnaire and the verbalization, we propose improvements of the IAS mainlyrelating to the views of State of the Traffic, State of the Line and Message interface agents.


Figure 5: Improvement given to both interfaces: State of the Traffic and State of the Line

6.1 Improvement relating to the view of State of the Traffic interface agent

For a better usability, we propose to introduce (in top of the window) a specific zone to thealarm, anomaly and warning messages. In fact, when a message appears, the user takes note ofit and validates it. Thus, instead of being lost, the message could be placed automatically in themessage planned zone. In this way, the user would not need more to memorize messages or tonote them on paper. This way of presenting the messages would make it possible the user to oneby one treat them and to remove them once treated.

6.2 Improvement relating to the view of State of the Line interface agent

To solve the problem related to the message sent to the stations and the vehicles, we proposeto introduce with the view of State of the Line interface agent a specific zone for the sending ofmessage (see Figure 5). This zone would make it possible the user to reach directly the Messageview without having to pass by the Station view or the Vehicle view.

6.3 Improvement relating to the Message interface agent

After the changes carried out on the view of State of the Traffic, the view of Messageinterface agent should be also changed. Indeed, if the user wishes to send a message to the wholeof the stations of a specific line, it would be interesting to mask the information relative to theunconcerned lines.


6.4 Other possible improvements

In addition, we note, during the evaluation, that the messages edited by the user are notrecorded by the IAS. Indeed, if the user wishes to send the same message twice, he or she has toreedit it.

7 Conclusion

We have presented in this article the results of a first evaluation of an Information Assis-tance System. This system is based on an agent-based architecture. This evaluation has beenperformed in laboratory and provided us with interesting results. The evaluation dealt with twodifferent scenarios. It was then possible to deduce from it first proposals for an improvement ofthe IAS.

Our further research aims, on the one hand, to the improvement of the Simulation mod-ules/Confrontation/Specification and generation of Petri nets (see [30], [29], [28]) and, on theother hand, a second evaluation on the ground which proves to be necessary for the detection ofthe utility and/or usability problems not detected during the first evaluation. The eye trackingtechnique [18] could be used to go deeper in the second evaluation.

8 Acknowledgements

The present research work has been supported by International Campus on Safety and Inter-modality in Transportation (CISIT), the Nord-Pas-de-Calais Region, the European Community(FEDER). The authors gratefully acknowledge the support of these institutions.

Bibliography

[1] L. Bass, R. Little, R. Pellegrino, and S. Reed. The arch model: Seeheim revisited. In UserInterface DeveloppersWorkshop, Seeheim, 1991.

[2] F. T. Bernard and B. David. AMF : un modèle d’architecture multi-agents multi-facettes.TSI, 18(5):555-586, 1999.

[3] J. Coutaz. PAC, on object oriented model for dialog design. In Interact’87, 1987. 6 pages.

[4] J. Coutaz and L. Nigay. Architecture logicielle conceptuelle des systèmes interactifs, pages207-246. 2001. Chapitre 7 Analyse et Conception de l’Interaction Homme-Machine dans lessyst‘emes d’information, Kolski Ed., Hermes Publ.

[5] I. Dzitac and B.E. Barbat. Artificial intelligence + distributed systems = agents. InternationalJournal of Computers Communications & Control, 1(4):17-26, 2009.

[6] H. Ezzedine, T. Bonte, C. Kolski, and C. Tahon. Integration of trafic management andtraveller information systems: basic principles and case study in intermodal transport systemmanagement. International Journal of Computers Communications & Control, 3(3):281-294,2008.

[7] H. Ezzedine and C. Kolski. Use of petri nets for modeling an agent-based interactive system:basic principles and case study. In V. Kordic (Ed.), Petri Net: Theory and Application,I-Tech Education and Publishing, pages 131-148, Vienna, Austria, 2008.


[8] H. Ezzedine, C. Kolski, and A. Péninou. Agent-oriented design of human-computer interface:application to supervision of an urban transport network. Eng. Appl. Artif. Intell., 18(3):255-270, April 2005.

[9] H. Ezzedine, A. Trabelsi, and C. Kolski. Modelling of an interactive system with an agent-based architecture using petri nets, application of the method to the supervision of a transportsystem. Mathematics and Computers in Simulation, 70(5-6):358-376, 2006.

[10] F. G. Filip and K. Leivisk¨a. Large-scale complex systems. In Handbook of Automation,pages 619-638. 2009.

[11] A. Goldberg. SMALLTALK-80: the interactive programming environment. Addison-WesleyLongman Publishing Co., Inc., Boston, MA, USA, 1984.

[12] D. Gracanin, P. Srinivasan, and K. P. Valavanis. Parameterized petri nets and their appli-cation to planning and coordination in intelligent systems. Systems, Man and Cybernetics,IEEE Transactions on, 24(10):1483-1497, oct 1994.

[13] A. Holyer. Methods for evaluating user interfaces. In School of Cognitive and ComputingSciences, University of Sussex, Brighton., 1993.

[14] M.Y. Ivory and M.A. Hearst. The state of the art in automating usability evaluation of userinterfaces. ACM Comput. Surv., 33(4):470-516, 2001.

[15] C. Kolski, P. Forbrig, B. David, P. Girard, C. D. Tran, and H. Ezzedine. Agent-basedarchitecture for interactive system design: Current approaches, perspectives and evaluation.In HCI (1), pages 624-633, 2009.

[16] J. Nielsen. Usability engineering. Academic Press, 1993.

[17] J. Nielsen and T. K. Landaue. A mathematical model of the finding of usability problems. InProceedings of the INTERACT ’93 and CHI ’93 conference on Human factors in computingsystems, CHI ’93, pages 206-213, New York, NY, USA, 1993. ACM.

[18] M. Pivec, C. Trummer, and J. Pripfl. Eye-tracking adaptable e-learning and content author-ing support. Informatica (Slovenia), 30(1):83-86, 2006.

[19] A. Sears and J.A. Jacko, editors. The Human Computer Interaction Handbook: Funda-mentals, Evolving Technologies and Emerging Applications. Lawrence Erlbaum Associates,Mahwah, NJ, 2. edition, 2008.

[20] S. Senach. Évaluation ergonomique des interfaces homme-machine: une revue de la littéra-ture. Technical report, Sophia Antipolis, 1990.

[21] R. Susannah and J. Graham. Evaluating usability of human-computer interfaces: a practicalmethod. Halsted Press, New York, NY, USA, 1989.

[22] J.C. Tarby, H. Ezzedine, and C. Kolski. Trace-based usability evaluation using aspect ori-ented programming and agent-based software architecture. In Human-Centered Software En-gineering, pages 257-276. 2009.

[23] A. Trabelsi. Contribution à l’évaluation des systèmes interactifs orientés agents : applicationà un poste de supervision du transport urbain. PhD thesis, University of Valenciennes andHainaut-Cambrésis, September 2006.


[24] A. Trabelsi and H. Ezzedine. Un pas vers un outil d’aide aux évaluateurs de systèmesinteractifs à base d’agents. In Colloque scientifique sur l’ergonomie et l’informatique avancée,pages 337-341, Biarritz, France, 2006.

[25] A. Trabelsi, H. Ezzedine, and C. Kolski. Architecture modelling and evaluation of agent-based interactive systems. In SMC (6), pages 5159-5164, 2004.

[26] A. Trabelsi, H. Ezzedine, and C. Kolski. Un mouchard électronique orienté agent pourl’évaluation de systèmes interactifs de supervision. In Conférence Internationale Francophoned’Automatique, Bordeaux, France, 2006.

[27] A. Trabelsi, H. Ezzedine, and C. Kolski. Evaluation of agent-based interactive systems,application to an information assistance system: first results. In European Annual Conferenceon Human Decision-Making and Manual Control, Reims, France, 2009.

[28] C. D. Tran. Vers un environnement genérique et configurable pour l’aide à l’évaluationdes systèmes interactifs à base d’agents, Application à un Système d’Aide à l’Informationvoyageurs. PhD thesis, University of Valenciennes and Hainaut-Cambrésis, July 2009.

[29] C. D. Tran, H. Ezzedine, and C. Kolski. Evaluation of agent-based interactive systems:Proposal of an electronic informer using petri nets. J. UCS, 14(19):3202-3216, 2008.

[30] C. D. Tran, H. Ezzedine, and C. Kolski. A generic and configurable electronic informer toassist the evaluation of agent-based interactive systems. In CADUI, pages 251-263, 2008.

[31] R. A. Virzi. Streamlining the design process: Running fewer subjects. In Human FactorsSociety 34th Annual Meeting, pages 291-294, 1993.

Author index

Borza S., 252Brandabur R.E., 183

Donoso Y., 207Duraiswamy K., 235Dzitac S., 215

Ezzedine H., 260

Karic M., 191

Lonea A.M., 223

Martinovic G., 191Marza B., 252Moisil I., 215Montoya G.A., 207

Pitic A.E., 215Popescu D.E., 223Prabhu P., 235

Serbu R., 252

Trabelsi A., 260

Velásquez-Villada C., 207