资源描述
,*,Intelligent Database Systems Lab,按一下以編輯母片標題樣式,按一下以編輯母片,第二層,第三層,第四層,第五層,N.Y.U.S.T.,I.M.,按一下以編輯母片標題樣式,按一下以編輯母片,第二層,第三層,第四層,第五層,*,Intelligent Database Systems Lab,EvolvingReactiveNPCs fortheReal-Time SimulationGame,Advisor,：,Dr.Hsu,Reporter,：,Wen-Hsiang Hu,Author,：,JinHyuk Hong and Sung-Bae Cho,IEEE SymposiumonComputational IntelligenceandGames,1,Outline,Motivation,Objective,Introduction,Thegame:Build&Build,Basicbehaviormodel,Co-evolutionarybehaviorgeneration,Experiment andResults,Discussion,Conclusion,PersonalOpinion,2,Motivation,AIincomputergameshasbeenhighlightedinrecent,but,manualworks,fordesigning theAI,cost agreatdeal,.,3,Objective,DesigningNPCsbehaviors withoutrelyingonhumanexpertise.,4,Basicbehaviormodel,Twodifferent gridscales areused fortheinputof,theneural networksuchas55and 1111.,five neuralnetworks,areusedtodecidewhethertheassociating,action,executesornot.,Thegame:Build&Build,randomaction probability:0.2,5,Co-evolutionarybehaviorgeneration,Weusethe,genetic algorithm,togeneratebehaviorsystems thatare accommodatedtoseveral environments.,6,Experiment andResults,55obtainslowerwinning averages forcomplex environment,whileitperformsbetterwhentheenvironmentisrather simple.,7,Introduction,Itischallengeable formany researchers to apply AI to controlcharacters.(AI producemorecomplex andrealisticgames.),Finitestatemachines,and,rule-based systems,arethe mostpopulartechniques in,designingthe movement of characters,.,While,neuralnetworks,Bayesiannetwork,and,artificial life,arerecentlyadoptedfor,flexiblebehaviors,.,Evolution,generatesuseful strategies,automatically,.,This paper proposes a,reactivebehaviorsystem,composedof,neuralnetworks,ispresented,andthe,systemisoptimized byco-evolution,.,8,Rule based approach,AIofmany computer games is designed with,rulesbasedtechniques,such as,finitestatemachines(FSMs),or,fuzzylogic,.,FSMs haveaweak point of itsstiffness;however,the,movement,ofa characterisapttobe,unrealistic,.,thereisa trend towardsfuzzystatemachine,(FuSM).,9,Adaptation andlearning:NNs,EAs,andArtificial life,The,adaptation,and,learning,ingameswill be oneofthemostmajorissues,makinggamesmoreinterestingandrealistic,.,Neuralnetwork,and,evolutionaryalgorithms,(e.g.genetic algorithm)are promisingartificial intelligencetechniques forlearningincomputergames.,NN-is badly trained,GE-required toomany computationsandweretooslowtoproduce usefulresults.,10,Co-evolution,Bysimultaneouslyevolvingtwoormorespecies withcoupledfitness.,Superiorstrategies foranenvironmenthave beendiscoveredbyco-evolutionaryapproaches.,11,Reactivebehavior,Reactivemodelperformseffectivelysinceitconsidersthe currentsituation only.,Neuralnetworksand behavior-based approachesare recently usedfor thereactivebehaviorofNPCs keepingthe realityofbehaviors.,12,Thegame:Build&Build,Build&Builddeveloped in thisresearchisareal-timestrategic simulationgame,inwhich,twonationsexpandtheirown territory,.,Each nationhassoldierswho individuallybuildtownsandfightagainsttheenemies,whileatown continually produces soldiers fora given period.,13,Thegame:Build&Build,14,Designingthe gameenvironment,Thegamestarts,twocompetitiveunits,ina restrictedland,withaninitialfund,.,Theunitsareabletotakesomeactionsatthenormallandbutnotattherockland.,Aunitcanbuildatownwhenthenationhasenoughmoney,whiletownsproduceunitsusingsomemoney.,15,Designingthegameenvironment,(cont.),16,DesigningNPCs,NPCcanmoveby4directions,aswellasbuildtowns,attackunitsortowns,and,mergewithotherNPCs,.,TheattackactionsareautomaticallyexecutedwhenanopponentlocatesbesidetheNPC.,17,DesigningNPCs,(cont.),18,DesigningNPCs,(cont.),19,Basicbehaviormodel(cont.),Twodifferentgridscalesareusedfortheinputoftheneuralnetworksuchas5,5and1111.,20,Basicbehaviormodel(cont.),Inordertoactivelyseeka,dynamicsituation,themodelselectsarandomactionwith,a,probability(inthispaper,a,=0.2)inadvance.,fiveneuralnetworks,areusedtodecidewhethertheassociating,action,executesornot.,21,Co-evolutionarybehaviorgeneration,Weusethe,geneticalgorithm,togeneratebehaviorsystemsthatareaccommodatedtoseveralenvironments.,Twopair-wisecompetitionpatternsareadoptedtoeffectivelycalculatethefitnessofanindividual.,22,Co-evolutionarybehaviorgeneration,(cont.),Thefitnessofanindividualismeasuredbythescoresagainstrandomlyselected,M,opponents.,23,ExperimentandResults,Fourdifferentbattle,maps=demonstrate,theproposedmethod,ingeneratingstrategies,adaptivetoeach,environment.,24,ExperimentandResults,(cont.),Thecasewith,1111 showsmorediversebehaviors,thanthat with 5,5,since itobserves information ona morelarge area.,55obtainslower winning averages for complex environment,while itperforms betterwhen the environment israthersimple.,25,Experiment and Results,(cont.),Fig.8.Winningratebetween55behavior and 11,11behaviorateachgenerationon map type3.,The1111 showsthebetter performance thanthe55,since it considersmorevariousinput conditions soasto generatediverseactions.,26,Experiment and Results,(cont.),Fortheplain map,5,5 behaviorsystem showsa,simple strategy,thattries to,build atown
展开阅读全文