设计全连接神经网络的权重分配算法(IJISA-V10-N6-8)

I.J.IntelligentSystemsandApplications,2018,6,68-76PublishedOnlineJune2018inMECS()DOI:10.5815/ijisa.2018.06.08Copyright©2018MECSI.J.IntelligentSystemsandApplications,2018,6,68-76WeightAssignmentAlgorithmsforDesigningFullyConnectedNeuralNetworkMrs.AartiM.KarandeResearchScholar,ComputerEngineering,S.P.I.T.MumbaiE-mail:[e-mail](IndustryRelation),HODComputerEngineering,S.P.I.T.MumbaiE-mail:[e-mail]:24May2017;Accepted:15September2017;Published:08June2018Abstract—Softcomputingisusedtosolvetheproblemswhereinputdataisincompleteorimprecise.Thispaperdemonstratedesigningfullyconnectedneuralnetworksystemusingfourdifferentweightcalculationalgorithms.Inputdataforweightcalculationisconstructedinthematrixformatbasedonthepairwisecomparisonofinputconstraints.Thiscomparisonisperformedusingsaaty’smethod.Thisinputmatrixhelpstobuildjudgmentbetweenseveralindividuals,formingasinglejudgment.AlgorithmconsideredhereareGeometricaveragemean,Linearalgebracalculation,Successivematrixsquaringmethod,andanalyticalhierarchicalprocessingmethod.Basedonthequalityparameterofperformance,itisobservedthatanalyticalhierarchicalprocessingisthemostpromisingmathematicalmethodforfindingappropriateweight.Analyticalhierarchicalprocessingworksonstructurationoftheproblemintosubproblems,HenceitthemostprominentmethodforweightcalculationinfullyconnectedNN.IndexTerms—SoftComputing,NeuralNetwork,Saaty’sMethod,AnalyticalHierarchicalProcessing,ExactLinearAlgebraCalculation,GeometricAverageApproximation,SuccessiveMatrixSquaring.I.INTRODUCTIONFindingasolutiontoarealtimeproblemisdifficultasitmaydependonthetypeofproblem,sensitivityoftheproblem,andtypeofsolutionexpected.Softcomputingapproachhelpstofindasolutioninunpredictablesituation.Neuralnetwork(NN)isoneofthetechniquesofthesofttechniques.NNrequiresweightstobeassignedamongneuronsforcalculatingresult.Thispaperdiscusses4differentalgorithmsforcalculatingweightsbasedoninputgiventotheneurons.Thispaperalsocomparesstatedalgorithmwithanexample.A.SoftComputingApproachUsingcomputertechnology,aprocesswhichcompletesataskiscalledascomputing.Computingisclassifiedintotwotypes.FirstisHardcomputing,i.e.,conventionalcomputingconceptwhichusestheavailableanalyticalcalculationmodel.SecondisSoftcomputing,whichworksasatolerantofimprecision,uncertainty,partialtruth,andapproximation.[28]Itdescribesandtransformsinformationsuchastheory,analysis,design,efficiency,implementation,andapplicationasthatofthehumanmind.[1][25]Themaincharacteristicofsoftcomputingisitsintrinsiccapabilitytocreatehybridsystemsbasedona(looseortight)integrationoftheconstituenttechnologies.[5]Thisintegrationcombinesuncleardomainknowledgeandempiricaldatatodevelopflexiblecomputingtoolsandtosolvecomplexproblemsbasedonthereasoning.[25][8][16]Softcomputingismoreorientedtowardstheanalysisanddesignofintelligentsystems.Softcomputing(SC)hasadifferenttechniquelikeneuralnetworks,fuzzylogic,geneticalgorithm,swarmoptimisationandtheirhybridcombination.[9]SChasstronglearningandcognitiveabilityandgoodtoleranceofuncertaintyandimprecision.SCtechniquesderivetheirpowerofgeneralisationgeneratingoutputfrompreviouslyunseenorlearnedinputsbasedonapproximation.[25]SoftComputingalgorithmsareusedforrealtimeunderstandableapplication.AdvantagesofSoftComputingcanbelistedasNonlinearproblemscanbesolvedusingSC;Itworksinhumanknowledgeareassuchascognition,recognition,understanding,learning,andcomputing.Intelligentsystemssuchasautonomousself-tuningsystems,andautomateddesignedsystemscanbeconstructedusingSC.II.STRUCTUREOFTHENEURALNETWORKA.ArtificialNeuralNetworkTheartificialneuralnetworkisoneofthemaintechniquesofsoftcomputing.[20][23][5]NNarenonlinearpredictivemodelsthatlearnthroughexperienceWeightAssignmentAlgorithmsforDesigningFullyConnectedNeuralNetwork69Copyright©2018MECSI.J.IntelligentSystemsandApplications,2018,6,68-76andtraining.NNmodelsarenon-parametricmodels.[17]NNareuniversalfunctionapproximators.[19]NNandstatisticalapproachesdifferinassumptionsaboutthestatisticaldistributionorpropertiesofdata.NNsgivesmuchaccuratewhenmodellingcomplexdata.NNsystemwilladapttotheinputenvironment.[12]Itwillproducetheoutputdependinguponitsinput.Theoutputofonenodecanbetheinputofanothernode.Thefinaloutputdependsonthecomplexinteractionamongallthenodes.[32][5]AcharacteristicofNNdependsonitsstructure,theconnectionstrengthbetweenneurons(i.e.synapticweightsbetweenneurons),i/pando/pnodeproperties,anderrorcalculationrules.Thepowerofneuralcomputationsdependsonconnectingneuronsinthenetwork.[30]NNsarepotentiallyusefulforunderstandingthecomplexrelationshipsbetweeninputandoutput.NNcanperformdataanalysisforimpreciseandincompleteinput.Theworkingofaneuralnetworkisdeterminedbyitsactivationfunctions,learningrule,andbasedontheconnectionofneuronsitself.[26]Duringtraining,theinter-unitconnectionsareoptimisedtilltheerrorinpredictionsisminimisedorthenetworkreachesthespecifiedlevelofaccuracy.Thenetworkistrainedandtestedwithinputpattern.Therearethreemajorstepsofex