折扣和平均準(zhǔn)則下SMDP基于性能勢(shì)的統(tǒng)一強(qiáng)化學(xué)習(xí)算法.pdf

1、分類號(hào)密級(jí)：一一單位代碼：]Q塹一9學(xué)號(hào)：03_1旦Q5280663侖肥工學(xué)火警HefeiUniversityofTechnology碩士學(xué)位論文MASTERDISSERTATIoN論文題目：學(xué)位類別：學(xué)科專業(yè)：(工程領(lǐng)域作者姓名：折扣和平均準(zhǔn)則下SMDP基于性能勢(shì)的統(tǒng)一強(qiáng)化學(xué)習(xí)算法學(xué)歷碩士計(jì)算機(jī)應(yīng)用技術(shù)周雷導(dǎo)師姓名_墮墨型墼堡完成時(shí)間：2006年5月UnifiedAlgorithmsforSemiMarkovDecisionProce

2、sseswithDiscountedandAverageCriteriaBasedonPerformancePotentialsbyReinforcementLearningAbstractAsaspecialtypeofrandomprocesses，MarkovprocesshasbeenusedinlotsoffieldsinouractuallifeMarkovdecisionprocesses(MDPs)andsemiMark

3、ovdecisionprocesses(SMDPs)arecommonlymathematicsmodelsfortheserandomsystemsSMDPsareaclassofmodel，whichismoregeneralthanMDPsThesojourntimeofitsstatesarefollowedasanarbitrarydistributionTheperformanceoptimizationofMDPsandS

4、MDPswithseveralcriteriaiSoneofthemostactiveresearchfieldsTheconceptofMarkovperformancepotentials，whichisintroducedbyCao，offersanewframeworkandapproachfortheoptimizationofMDPsBasedonoptimalPoissonequationandtheoptimaltheo

5、remwithpotentialsaloofalgorithms，suchaspolicyiterationandvalueiteration，canbeobtainedInrecentyears，reinforcementlearningismoreandmoreappliedintotheseliteratures。Itisanimportantfieldofartificialintelligence，whichcombinesc

6、onceptsfromdynamicprogramming，stochasticapproximationviasimulation，andfunctionapproximationIthasbeenshowntosupplyoptimalornearoptimalsolutionstolargeMDPs／SMDPsevenwhentheyareconsideredintractableviatraditionaldynamicprog

7、rammingmethods。Accordingtotherelationsbetweenperformancepotentialsandreinforcementlearning，thepaperisconcernedwiththeperformanceoptimizationalgorithmsforSMDPsbysimulation，basedapproximationofpotentialsFirst，bythedefiniti

8、onofanequivalentinfinitesimalgeneratorofSMDPs，SMDPscanbeconvertedintoanequivalentuniformizedMarkovchainSomeconclusionsforMDPsmaybeextendedintoSMDPs，Second，bythedefinitionofpotentialsbasedonPoissonequationandsamplepath，th

9、euniformapproximationformulaofpotentialscanbeobtainedwiththediscountedandaveragecriteriabyreinforcementlearning。Further，anewapproach，neuro—dynamicprogramming(NDP)，whichcansolvetheoptimizationproblemoflarge—scalediscretee