元博弈平衡和多Agent强化学习的MetaQ算法.pdfVIP

计算机研究与发展 ISSN1000．1239|CN11—1777|TP of Researchand JournalComputer Development 元博弈平衡和多Agent强化学习的MetaQ算法 王 皓 高 阳 (南京大学软件新技术国家重点实验室南京210093) (1eafwanghao@gmail．conl) and for Reinforcement Meta-EquilibriaMetaQAlgorithmMulti-Agent Learning HaoandGao Wang Yang Novel (NationalLaboratoryfor SoftwareTechnology，NanfingUniversity，Nanjing210093) anextension Abstract reinforcement ofreinforcement Multiagent learning(MARL)is learning(RL) from to domain．Nash is mentionedasa in singleagentmultiagent Q—Learningfrequently pivotalalgorithm arethreemaindefectsin useofmixed looks multiagentsystem．However，there NashQ．First，thestrategy somewhat Nash ofa could beParetodominated．And ambiguous．Second，theequilibriumgame probably the costofNash be defectscomefromthe third，thecomputationalequilibriamayexpensive．All hypothesis that areNash basedon is agents rational．Thus，a“MetaQ”algorithmmetagametheoryproposed．MetaQ uses thanNash avoidallthedefects．Theoreticand rationality，to meta．rationalily，rather experimental madeabout 1cadstotheconclusionthat hasabetter than studiesare MetaQ，which MetaQ performance NashQ． wordsreinforcement Key learning；multiagentsystem；metagame；MetaQ 摘要 Q学习算法以元博弈理论 足都来源于“Agent是Nash理性的”这一假设．一个称为“MetaQ”的多Agent 为基础，通过改变Agent的理性来避免所有的这些不足．研究证