部分可觀察markov決策過程中基于內(nèi)部狀態(tài)的強化學(xué)習(xí)研究

1、合肥工業(yè)大學(xué)碩士學(xué)位論文部分可觀察Markov決策過程中基于內(nèi)部狀態(tài)的強化學(xué)習(xí)研究姓名：方長勝申請學(xué)位級別：碩士專業(yè)：計算機應(yīng)用技術(shù)指導(dǎo)教師：王浩20080601TheReinforcementLearningResearchBasedOnInternalStateInPartiallyObservableMarkovDecisionProcessesAbstractReinfo羚ememlearning(RE)isallimporta

2、ntbranchofmachinelearningRLdevisesthemappolicyfromstatestoactionsby”trial—error’’principleandleamsteactunderallstates，SOthattheadaptabilityandrobustnessofAIsystemscouldbeimprovedInspireofsomeachievementinthisareathere掰℃s

3、tillmanyproblemsunsolved，partiallyo鶼娃vabkissueisoneofthem。ThePOMDPisallidealmodeltotacklesuchkindofproblemsHoweverthemodelingPOMDPcomesatapriceexactmetho&forSOlvingthemalecomputationallyveryexpensiveandthusapplicableinpr

4、acticeonlytoverysimpleproblemsWedoSOmeworksaboutmodeloptimizingandalgorithmimprovingoninthisdissertationThemainworksaleasfollowsFirstlyinthePOMDPmodel，theinternalstateofagentisintroducedandtheexperienceofagentisused。TheP

5、OMDPreinfo＆ementlearningimprovedmodelbasedoninternalstateisproposedinthisdissertation。ThedescriptionoftheexamplepolicyindicatesthatthepolicycomplexityreducedandthelearningefficiencyimprovedSecondlytheeligibilitytraceisin

6、troducedintothemodelThePGI—POMDPalgorithm，approximatereinforcementlearningalgorithmbasedonpolicygradientmethods，isproposed。TheresultshaveprovedthatPGIPOMDPalgorithmcarlreducethecomputablecomplexityandimprovethecomputatio

7、nefficiencyThirdlythemethodisappliedonMAS，andMIS—GPOMDPalgorithmisproposedTheMISGPOMDPalgorithmisoneofthepolicy伊adiemmethodsonMASTheexperimentalresultshaveshownthatthelearningefficiencyandthecostoftimeandspacealebothimpr