,*,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,Evolving Reactive NPCs for the Real-Time Simulation Game,Advisor,：,Dr.Hsu,Reporter,：,Wen-Hsiang Hu,Author,：,JinHyuk Hong and Sung-Bae Cho,IEEE Symposium on Computational Intelligence and Games,1,Outline,Motivation,Objective,Introduction,The game:Build&Build,Basic behavior model,Co-evolutionary behavior generation,Experiment and Results,Discussion,Conclusion,Personal Opinion,2,Motivation,AI in computer games has been highlighted in recent,but,manual works,for designing the AI,cost a great deal,.,3,Objective,Designing NPCs behaviors without relying on human expertise.,4,Basic behavior model,Two different grid scales are used for the input of,the neural network such as 55 and 1111.,five neural networks,are used to decide whether the associating,action,executes or not.,The game:Build&Build,random action probability:0.2,5,Co-evolutionary behavior generation,We use the,genetic algorithm,to generate behavior systems that are accommodated to several environments.,6,Experiment and Results,55 obtains lower winning averages for complex environment,while it performs better when the environment is rather simple.,7,Introduction,It is challengeable for many researchers to apply AI to control characters.(AI produce more complex and realistic games.),Finite state machines,and,rule-based systems,are the most popular techniques in,designing the movement of characters,.,While,neural networks,Bayesian network,and,artificial life,are recently adopted for,flexible behaviors,.,Evolution,generates useful strategies,automatically,.,This paper proposes a,reactive behavior system,composed of,neural networks,is presented,and the,system is optimized by,co-evolution,.,8,Rule based approach,AI of many computer games is designed with,rules based techniques,such as,finite state machines(FSMs),or,fuzzy logic,.,FSMs have a weak point of its stiffness;however,the,movement,of a character is apt to be,unrealistic,.,there is a trend towards fuzzy state machine,(FuSM).,9,Adaptation and learning:NNs,EAs,and Artificial life,The,adaptation,and,learning,in games will be one of the most major issues,making games more interesting and realistic,.,Neural network,and,evolutionary algorithms,(e.g.genetic algorithm)are promising artificial intelligence techniques for learning in computer games.,NN-is badly trained,GE-required too many computations and were too slow to produce useful results.,10,Co-evolution,Bysimultaneouslyevolvingtwoormorespecieswithcoupledfitness.,Superiorstrategiesforanenvironmenthavebeendiscoveredbyco-evolutionaryapproaches.,11,Reactivebehavior,Reactivemodelperformseffectivelysinceitconsidersthecurrentsituationonly.,Neuralnetworksandbehavior-basedapproachesarerecentlyusedforthereactivebehaviorofNPCskeepingtherealityofbehaviors.,12,Thegame:Build&Build,Build&Build,developedinthisresearchisareal-timestrategicsimulationgame,inwhich,twonationsexpandtheirownterritory,.,Eachnationhassoldierswhoindividuallybuildtownsandfightagainsttheenemies,whileatowncontinuallyproducessoldiersforagivenperiod.,13,Thegame:Build&Build,14,Designingthegameenvironment,Thegamestarts,twocompetitiveunits,inarestrictedland,withaninitialfund,.,Theunitsareabletotakesomeactionsatthenormallandbutnotattherockland.,Aunitcanbuildatownwhenthenationhasenoughmoney,whiletownsproduceunitsusingsomemoney.,15,Designingthegameenvironment,(cont.),16,DesigningNPCs,NPCcan moveby4directions,aswell as build towns,attackunitsortowns,and,mergewith other NPCs,.,Theattack actionsare automaticallyexecutedwhenanopponentlocates besidetheNPC.,17,DesigningNPCs,(cont.),18,DesigningNPCs,(cont.),19,Basicbehaviormodel(cont.),Twodifferent gridscales areused fortheinputofthe neuralnetwork suchas55 and1111.,20,Basicbehaviormodel(cont.),Inordertoactivelyseek a,dynamic situation,themodelselects arandom actionwith,a,probability(inthispaper,a,=0.2)inadvance.,five neuralnetworks,areusedtodecidewhethertheassociating,action,executesornot.,21,Co-evolutionarybehaviorgeneration,Weusethe,genetic algorithm,togeneratebehaviorsystems thatare accommodatedtoseveral environments.,Twopair-wise competition patterns areadopted to effectively calculatethefitnessofanindividual.,22,Co-evolutionarybehaviorgeneration,(cont.),Thefitnessofanindividual is measured by thescoresagainstrandomlyselected,M,opponents.,23,Experiment andResults,Four differentbattle,maps=demonstrate,theproposedmethod,ingenerating strategies,adaptivetoeach,environment.,24,Experiment andResults,(cont.),Thecasewith,1111showsmore diversebehaviors,than thatwith55,since it observes information on amorelargearea.,55obtainslowerwinning averages forcomplex environment,whileitperformsbetterwhentheenvironmentisrather simple.,25,Experiment andResults,(cont.),Fig.8.Winningratebetween 5,5behaviorand11,11behaviorateachgeneration on maptype 3.,The11,11showsthe betterperformancethan the55,since it considersmore vari