Mastering the game of Go with deep neural networks and tree research.pdf

《Mastering the game of Go with deep neural networks and tree research.pdf》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、ARTICLEdoi:10.1038/nature16961MasteringthegameofGowithdeepneuralnetworksandtreesearchDavidSilver1*,AjaHuang1*,ChrisJ.Maddison1,ArthurGuez1,LaurentSifre1,GeorgevandenDriessche1,JulianSchrittwieser1,IoannisAntonoglou1,VedaPanneershelvam1,MarcLanctot1,SanderDieleman1,DominikGrewe1,JohnNham2,NalKal

2、chbrenner1,IlyaSutskever2,TimothyLillicrap1,MadeleineLeach1,KorayKavukcuoglu1,ThoreGraepel1&DemisHassabis1ThegameofGohaslongbeenviewedasthemostchallengingofclassicgamesforartificialintelligenceowingtoitsenormoussearchspaceandthedifficultyofevaluatingboardpositionsandmoves.Hereweintroduceanewapp

3、roachtocomputerGothatuses‘valuenetworks’toevaluateboardpositionsand‘policynetworks’toselectmoves.Thesedeepneuralnetworksaretrainedbyanovelcombinationofsupervisedlearningfromhumanexpertgames,andreinforcementlearningfromgamesofself-play.Withoutanylookaheadsearch,theneuralnetworksplayGoatthelevelo

4、fstate-of-the-artMonteCarlotreesearchprogramsthatsimulatethousandsofrandomgamesofself-play.WealsointroduceanewsearchalgorithmthatcombinesMonteCarlosimulationwithvalueandpolicynetworks.Usingthissearchalgorithm,ourprogramAlphaGoachieveda99.8%winningrateagainstotherGoprograms,anddefeatedthehumanEu

5、ropeanGochampionby5gamesto0.Thisisthefirsttimethatacomputerprogramhasdefeatedahumanprofessionalplayerinthefull-sizedgameofGo,afeatpreviouslythoughttobeatleastadecadeaway.Allgamesofperfectinformationhaveanoptimalvaluefunction,v*(s),policies13–15orvaluefunctions16basedonalinearcombinationofwhichd

6、eterminestheoutcomeofthegame,fromeveryboardpositioninputfeatures.orstates,underperfectplaybyallplayers.ThesegamesmaybesolvedRecently,deepconvolutionalneuralnetworkshaveachievedunprec-byrecursivelycomputingtheoptimalvaluefunctioninasearchtreeedentedperformanceinvisualdomains:forexample,imageclas

7、sifica-containingapproximatelybdpossiblesequencesofmoves,wherebistion17,facerecognition18,andplayingAtarigames19.Theyusemanythegame’sbreadth(numberoflegalmovesperposition)anddisitslayersofneurons,eacharrangedinoverlappingtiles,toc