游戏要好玩需发展良好的AI

IntelligentDatabaseSystemsLab國立雲林科技大學NationalYunlinUniversityofScienceandTechnology1EvolvingReactiveNPCsfortheReal-TimeSimulationGameAdvisor：Dr.HsuReporter：Wen-HsiangHuAuthor：JinHyukHongandSung-BaeChoIEEESymposiumonComputationalIntelligenceandGamesIntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.2OutlineMotivationObjectiveIntroductionThegame:Build&BuildBasicbehaviormodelCo-evolutionarybehaviorgenerationExperimentandResultsDiscussionConclusionPersonalOpinionIntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.3MotivationAIincomputergameshasbeenhighlightedinrecent,butmanualworksfordesigningtheAIcostagreatdeal.IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.4ObjectiveDesigningNPCs’behaviorswithoutrelyingonhumanexpertise.IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.5BasicbehaviormodelTwodifferentgridscalesareusedfortheinputoftheneuralnetworksuchas5×5and11×11.fiveneuralnetworksareusedtodecidewhethertheassociatingactionexecutesornot.Thegame:Build&Buildrandomactionprobability:0.2IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.6Co-evolutionarybehaviorgenerationWeusethegeneticalgorithmtogeneratebehaviorsystemsthatareaccommodatedtoseveralenvironments.IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.7ExperimentandResults5×5obtainslowerwinningaveragesforcomplexenvironment,whileitperformsbetterwhentheenvironmentisrathersimple.IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.8IntroductionItischallengeableformanyresearcherstoapplyAItocontrolcharacters.(AIproducemorecomplexandrealisticgames.)Finitestatemachinesandrule-basedsystemsarethemostpopulartechniquesindesigningthemovementofcharacters.Whileneuralnetworks,Bayesiannetwork,andartificiallifearerecentlyadoptedforflexiblebehaviors.Evolutiongeneratesusefulstrategiesautomatically.Thispaperproposesareactivebehaviorsystemcomposedofneuralnetworksispresented,andthesystemisoptimizedbyco-evolution.IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.9RulebasedapproachAIofmanycomputergamesisdesignedwithrulesbasedtechniquessuchasfinitestatemachines(FSMs)orfuzzylogic.FSMshaveaweakpointofitsstiffness;however,themovementofacharacterisapttobeunrealistic.─thereisatrendtowardsfuzzystatemachine(FuSM).IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.10Adaptationandlearning:NNs,EAs,andArtificiallifeTheadaptationandlearningingameswillbeoneofthemostmajorissuesmakinggamesmoreinterestingandrealistic.Neuralnetwork,andevolutionaryalgorithms(e.g.geneticalgorithm)arepromisingartificialintelligencetechniquesforlearningincomputergames.─NN-isbadlytrained─GE-requiredtoomanycomputationsandweretooslowtoproduceusefulresults.IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.11Co-evolutionBysimultaneouslyevolvingtwoormorespecieswithcoupledfitness.Superiorstrategiesforanenvironmenthavebeendiscoveredbyco-evolutionaryapproaches.IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.12ReactivebehaviorReactivemodelperformseffectivelysinceitconsidersthecurrentsituationonly.Neuralnetworksandbehavior-basedapproachesarerecentlyusedforthereactivebehaviorofNPCskeepingtherealityofbehaviors.IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.13Thegame:Build&Build‘Build&Build’developedinthisresearchisareal-timestrategicsimulationgame,inwhichtwonationsexpandtheirownterritory.Eachnationhassoldierswhoindividuallybuildtownsandfightagainsttheenemies,whileatowncontinuallyproducessoldiersforagivenperiod.IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.14Thegame:Build&BuildIntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.15DesigningthegameenvironmentThegamestartstwocompetitiveunitsinarestrictedlandwithaninitialfund.Theunitsareabletotakesomeactionsatthenormallandbutnotattherockland.Aunitcanbuildatownwhenthenationhasenoughmoney,whiletownsproduceunitsusingsomemoney.IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.16Designingthegameenvironment(cont.)IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.17DesigningNPCsNPCcanmoveby4directionsaswellasbuildtowns,attackunitsortowns,andmergewithotherNPCs.TheattackactionsareautomaticallyexecutedwhenanopponentlocatesbesidetheNPC.IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.18DesigningNPCs(cont.)IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.19DesigningNPCs(cont.)IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.20Basicbehaviormodel(cont.)Twodifferentgridscalesareusedfortheinputoftheneuralnetworksuchas5×5and11×11.IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.21Basicbehaviormodel(cont.)Inordertoactivelyseekadynamicsituation,themodelselectsarandomactionwithaprobability(inthispaper,a=0.2)inadvance.fiveneuralnetworksareusedtodecidewhethertheassociatingactionexecutesornot.IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.22Co-evolutionarybehaviorgenerationWeusethegeneticalgorithmtogeneratebehaviorsystemsthatareaccommodatedtoseveralenvironments.Twopair-wisecompetitionpatternsareadoptedtoeffectivelycalculatethefitnessofanindividual.IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.23Co-evolutionarybehaviorgeneration(cont.)ThefitnessofanindividualismeasuredbythescoresagainstrandomlyselectedMopponents.IntelligentDatabaseSystemsLabN.Y.U.S.T.I.M.24ExperimentandResultsFourdifferentbattlemaps=demonstrate