成像技术对粒子群优化算法的推广对神经网络的影响(IJIGSP-V12-N2-2)

I.J.Image,GraphicsandSignalProcessing,2020,2,9-18PublishedOnlineApril2020inMECS()DOI:10.5815/ijigsp.2020.02.02Copyright©2020MECSI.J.Image,GraphicsandSignalProcessing,2020,2,9-18ExploringtheEffectofImagingTechniquesExtensiontoPSOonNeuralNetworksAnesA.AbbasDepartmentofComputerScience,UniversityofBahrainEmail:[e-mail]erScience,UniversityofBahrainEmail:[e-mail]:30August2019;Accepted:17November2019;Published:08April2020Abstract—Inthispaperwegothroughsomeveryrecentimagingtechniquesthatareinspiredfromspaceexploration.Theadvantagesofthesetechniquesaretohelpinsearchingspace.Toexploretheeffectivenessoftheseimagingtechniquesonsearchspaces,weconsidertheParticleSwarmOptimizationalgorithmandextenditusingtheimagingtechniquestotrainmultipleneuralnetworksusingseveraldatasetsforthepurposeofclassification.Thetechniqueswereusedduringthepopulationinitializationstageandduringthemainsearch.Theperformanceofthetechniqueshasbeenmeasuredbasedonvariousexperiments,thesetechniqueshavebeenevaluatedagainsteachother,andagainsttheparticleswarmoptimizationalgorithmalonetakingintoaccounttheclassificationaccuracyandtrainingruntime.Theresultsshowthattheuseofimagingtechniquesproducesbetterresults.IndexTerms—SearchSpaceImaging,Metaheuristics,Optimization,ParticleSwarmOptimization,ArtificialNeuralNetworks,PopulationInitialization.I.INTRODUCTIONThemainpurposeofallmetaheuristicalgorithmsistoexplorethesearchspaceandhelpinreachingtotheoptimumornear-optimumsolution.Someofthemainproblemsthatmightnothelpthemetaheuristictechniquestoreachtooptimumsolutionsiseitherbecauseofthecomplexityoftheproblemorproblemresourcesareunclearorlimited[15].Mostofthemetaheuristicalgorithmsareinspiredandadoptedfrombiologyornaturesuchaschromosomes,birds,fishswarmandbats[10].Innature,humanstrytoexplorespacethroughvarioustoolssuchastelescopes,satellitesandradars.Usuallyexpeditionstodiscoverthespacearesentafterverydeepinvestigationsthroughthepreviousmentionedtools.Thiswillhelpinlimitingthescopeofexplorationtowardsthescientist’starget.Theclosertheexpeditiontothetarget,themoreeffortwillbedonestartingfromthereachedposition.Thegeneralideaistomaintainrandomnessundercontrol.Mostofthemetaheuristictechniquesdependonrandomnessbutusuallyrandomnessmightleadtomorecostintermsoftimeandresources.Toreducethisproblem,humanexpeditionsmoveinacontrolledrandomnesstoensurereachingtonear-optimumsolutionconsumelesscost.Manyresearchersworktoeitherproposenewmetaheuristictechniquesinspiredfrombiologyordevelopsystemsthatcombinevariousmetaheuristictechniques.ProposingnewtechniquessuchasGeneticAlgorithm(GA)[11],SimulatedAnnealing(SA)[17],antcolony[6,7],batalgorithm[31],PSOandfishswarm[16][26],andCombiningmorethanonetechniquesuchasPSOandGA,GAandSA,orPSOandSA[4][12,13][27][31].Inboththecases,proposingnewmetaheuristicorcombinemorethanonetechnique,thetargetistobetterexplorethesearchspaceandachievebetterresults(near-optimum).In[22]RichardsandVenturaproposedatechniquetoinitializethepopulationcalledcentroidalVoronoitessellatiwhichstartswithapopulationthatisgeneratedrandomlytheniterativelytriesmaketheparticlesmovefarfromeachotheraspossibletohaveadiversityintheinitialpopulation.In[21]authorsproposedanothertechniqueforpopulationinitializationwheretherandomparticlesaregeneratedandwitheachgeneratedparticleitscomplement/inverseparticleisalsogenerated.Thenfromtheseparticlestheinitialpopulationisgenerated.In[18]Maaranenet.alproposedatechniquecalledquasi-randomgeneratortechniquetoinitializethepopulation.Thistechniqueformsrepetitivepatternstoavoidhavingmanyparticlesinsimilar/closelocations.Thistechnique’sideaisbasedongeneratingapopulationwithhighdiversityaspossible.Basedontheresearcherswork,theytrytohaveadiverseinitialpopulationwhichmeanshavingdiverseparticlessothattheycancapturethemostofthegoodparticles.However,thismightnotalwaystruebecausethismightbedoneregardlessoftheimportanceoftheseparticles.Theimagingtechniquedependbasicallyonexplorationfirstthenexpedition,whichmeanscheckingfirstthepotentialparticlesthenformingthepopulation.10ExploringtheEffectofImagingTechniquesExtensiontoPSOonNeuralNetworksCopyright©2020MECSI.J.Image,GraphicsandSignalProcessing,2020,2,9-18Fig.1.Particleswarmoptimizationpseudocode;pBest:Bestsolutionfoundbyparticle,gBest:Bestsolutionfoundbyswarm[28]Inthisresearch,ourobjectivesistoexploretheeffectofimagingtechniquesasextendtoPSOintheclassificationofneuralnetworks.Wearenotinterestedtocomparetheobtainedresultswithotherapproaches,butinsteadwecareaboutcomparingtheobtainedresultsusingextendPSOandtheregularPSO.Toevaluateimagingtechniques,theyareimplementedtoextendaparticleswarmoptimizationmetaheuristicusedtotunetheweightsandbiasesofmultipleclassificationneuralnetworksinsevendifferentdatasets.Thus,thefollowingthreesubsectionsbrieflyintroducetheparticleswarmoptimizationmetaheuristicandtheartificialneuralnetworktechniques.A.ParticleSwarmOptimizationPSOisaverywellknowmetaheuristicsearchalgorithmusedinvariousapplicationsrangi