Self Learning AI-Agents Part I_ Markov Decision Processes

《Self Learning AI-Agents Part I_ Markov Decision Processes》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、Applausefromyou,LudovicBenistant,and115othersArtemOppermannFollowMSc.Physics

2、DeepLearning&AISoftwareDeveloper

3、Expertisein#PredictiveAnalytics#Reinforcement Learning#ComputerVision#NLP#AIforBusinessOct 14·11minreadSelfLearningAI-AgentsPartI:MarkovDecisionProcessesAmathematicalguideonthetheorybehin

4、dDeepReinforcementLearningThisistherstarticleofthemulti-partseriesonselflearningAI-Agentsortocallitmoreprecisely—DeepReinforcementLearning.Theaimoftheseriesisn’tjusttogiveyouanintuitiononthesetopics.RatherIwanttoprovideyouwithmoreindepthcomprehensionofthetheory,mathematicsandimplementationbehind

5、themostpopularandeectivemethodsofDeepReinforcementLearning.SelfLearningAI-AgentsSeries—TableofContent•PartI:MarkovDecisionProcesses(Thisarticle)•PartII:DeepQ-Learning•PartIII:Deep(Double)Q-Learning•PartIV:PolicyGradientsforContinuesActionSpaces•PartV:DuelingNetworks•PartVI:AsynchronousActor-Crit

6、icAgents•…Fig.1.AIagentlearnedhowtorunandovercomeobstacles.MarkovDecisionProcesses—TableofContent•0.Introduction•1.ReinforcementLearninginaNutshell•2.MarkovDecisionProcesses•2.1MarkovProcesses•2.2MarkovRewardProcess•2.3ValueFunction•3.BellmanEquation•3.1BellmanEquationforMarkovRewardProcesses•3.2

7、MarkovDecisionProcess—Denition•3.3Policies•3.4Action-ValueFunction•3.5OptimalPolicy•3.6BellmanOptimalityEquation...0.IntroductionDeepreinforcementlearningisontherise.Noothersub-eldofDeepLearningwasmoretalkedaboutintherecentyears-bytheresearchersaswellasthemassmediaworldwide.Mostoutstandingachie

8、vementsindeeplearningweremadeduetodeepreinforcementlearning.FromGoogle’sAlphaGothathavebeatentheworldsbesthumanplayerintheboardgameGo(anachievementthatwasassumedimpossibleacoupleyearsprior)toDeepMind’sAIagentsthatteachthemselvestowalk,runandovercomeobstacles(Fig.1–3).Fig.2.AIagentlearnedhowtoruna

9、ndovercomeobstacles.Fig.3.AIagentlearnedhowtorunandovercomeobstacles.OtherAIagentsexceedsince2014humanlevelperformancesinplayingoldschoolAtarigamessuchasBreakthrough(Fig.4).Themostamazingthingaboutallofthisinmyopinioni