基于神经网络的机器学习在改进COCOMO模型中的应用(IJITCS-V10-N3-5)

I.J.InformationTechnologyandComputerScience,2018,3,35-51PublishedOnlineMarch2018inMECS()DOI:10.5815/ijitcs.2018.03.05Copyright©2018MECSI.J.InformationTechnologyandComputerScience,2018,3,35-51MachineLearningApplicationtoImproveCOCOMOModelusingNeuralNetworksSomyaGoyalVaishCollegeofEngineering/CSEDepartment,Rohtak,124001,INDIAE-mail:[e-mail]University/CSEDepartment,Jaipur,303007,INDIAE-mail:[e-mail]:13November2017;Accepted:08January2018;Published:08March2018Abstract—Millionsofcompaniesexpendbillionsofdollarsontrillionsofsoftwareforthedevelopmentandmaintenance.Stillmanyprojectsresultinfailurecausingheavyfinancialloss.Majorreasonistheinefficienteffortestimationtechniqueswhicharenotsosuitableforthecurrentdevelopmentmethods.Thecontinuouschangeinthesoftwaredevelopmenttechnologymakeseffortestimationmorechallenging.Tilldate,noestimationmethodhasbeenfoundfull-prooftoaccuratelypre-computethetime,money,effort(man-hours)andotherresourcesrequiredtosuccessfullycompletetheprojectresultingeitherover-estimatedbudgetorunder-estimatedbudget.HereamachinelearningCOCOMOisproposedwhichisanovelnon-algorithmicapproachtoeffortestimation.Thisestimationtechniqueperformswellwithintheirpre-specifieddomainsandbeyondso.Asdevelopmentmethodshaveundergonerevolutionariesbutestimationtechniquesarenotsomodifiedtocopeupwiththemoderndevelopmentskills,sotheneedoftrainingthemodelstoworkwithupdateddevelopmentmethodsisbeingsatiatedjustbyfindingoutthepatternsandassociationsamongthedomainspecificdatasetsvianeuralnetworksalongwithcarriageofdesiredCOCOMOfeatures.Thispaperestimatestheeffortbytrainingproposedneuralnetworkusingalreadypublisheddata-setandlateron,thetestingisdone.Thevalidationclearlyshowsthattheperformanceofalgorithmicmethodisimprovedbytheproposedmachinelearningmethod.IndexTerms—COCOMO(ConstructiveCostModel),Correlation,MachineLearning,MMRE(MeanMagnitudeofRelativeError),NeuralNetwork,SoftwareEffortEstimation.I.INTRODUCTIONEverydomainofourlifeiscoveredwithoverwhelmingapplicationsofcomputersnow-a-days.Noaspectleftuntouchedbycomputersi.e.hardwareandsoftware.Itisfoundthatthepricesforcomputerhardwarehasdecreasedincomparisonofsoftwarewhichiscontinuouslyincreasing.SoftwareIndustryannuallyspendthebillionsontheacquisitionandmaintenanceofsoftware[1].Object-orientedprogramming,computer-aidedsoftwareengineering(CASE),COTS,AgileMethodologyandothertechnologyareinuseforsoftwaredevelopment,butsoftwareeffortestimationhassomewherelaggedbehindintermsofadvancements.OnemajorresourceforsoftwareproductisMan-power,theeffort.Estimationmodelsfirstcomputetheeffortrequiredtocompletetheproject,thatcanbefurtherconvertedintodollars.Thecurrentestimationmodelsdisheartenprojectmanagersbyover-estimatedbudgetorunder-estimatedbudgetresultingintoacompletefailure.Variousmodelsforsoftwarecostestimationareavailableinmarket.ThemostpopularoneistheConstructiveCostModel,orCOCOMOdevelopedbyBarryBoehm[2].ThebasisforCOCOMOisadatabaseofsixty-threeprojectscreatedatTRWduringthe1960'sand1970'sandispublishedinBoehm'sbook,SoftwareEngineeringEconomics.ThepopularityofCOCOMOliesinitseaseofapplicationanditsnon-proprietarynature.Othermodels,likeESTIMACSareproprietary.Inallthesemodelstheinputsandtherelationshipsaredomainspecificwhicharefullydependentontheexpertsopinion.Forthisreason,suchmodelstendtoperformpoororevenfailwhentheirapplicationboundariesaretriedtobechanged.Insuchscenario,thereisaneedofatechniquethatcansubstitutetheexpert-judgement.ThedestinedanswerisMachineLearning.DataCollection,Knowledgeacquisition,classification,patternrecognitionandmuchmorecanbedoneeasilyandefficiently.HerewetriedtoapplyMachineLearningtoSoftwareEngineeringinEffortEstimation.NeuralNetworkallowstomodelacomplexsetofrelationshipbetweenthedependentvariableandtheindependentvariables.Ifweconsidereffortasdependentattributeandcostdriverswithsoftwaresizeasindependententitiesthen,neuralnetworkcanbeimplementedasmachinelearningtool.Theoverallobjectiveistodesignamethodologyformachinelearningbasedapproachtosoftwareeffortestimationusingneuralnetworks.Becausecurrentestimationmodelsprovideonlymarginalresultswithin36MachineLearningApplicationtoImproveCOCOMOModelusingNeuralNetworksCopyright©2018MECSI.J.InformationTechnologyandComputerScience,2018,3,35-51theirdomainspecificapplicationsotherwise,tendtofailwhenappliedforlatestdevelopmentmethods.Theperformanceofproposedmodelbasedonmachine-learningtechniqueisevaluatedandanalyzedincomparisontothetraditionalmodels.ThisproposedmodelusestheCOCOMOdatasetfortrainingphaseandtheKemmererdatasetfortestingpurposes.Then,theeffectivenessofmachine-learningtechniquetotheeffortestimationfieldisdeterminedandpost-analysisconclusionisdrawn.Withthisresearchwork,wetriednotonlytodevelopabetteropportunityforcostestimationbutalsotriedtoapplymachinelearningtosoftwareengineering.Thispaperisstructuredintosixsectionsasfollows:SectionII,discussesthebackgroundoftheresearchworktheme.SectionIII,coverstheliteraturesurveyhighlightingthevar