返回
Better Computer Go Player with Neural Network and Long-term Prediction.pdf

Better Computer Go Player with Neural Network and Long-term Prediction.pdf

ID：33943670

大小：1.59 MB

页数：10页

时间：2019-03-02

Better Computer Go Player with Neural Network and Long-term Prediction.pdf_第1页
Better Computer Go Player with Neural Network and Long-term Prediction.pdf_第2页
Better Computer Go Player with Neural Network and Long-term Prediction.pdf_第3页
Better Computer Go Player with Neural Network and Long-term Prediction.pdf_第4页
Better Computer Go Player with Neural Network and Long-term Prediction.pdf_第5页
资源描述：

《Better Computer Go Player with Neural Network and Long-term Prediction.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、UnderreviewasaconferencepaperatICLR2016BETTERCOMPUTERGOPLAYERWITHNEURALNET-WORKANDLONG-TERMPREDICTIONYuandongTianYanZhuFacebookAIResearchRutgersUniversityMenloPark,CA94025FacebookAIResearchyuandong@fb.comyz328@cs.rutgers.eduABSTRACTCompetingwithtophumanplayersintheancientgameofGohasbeenalong-termg

2、oalofartificialintelligence.Go’shighbranchingfactormakestraditionalsearchtechniquesineffective,evenonleading-edgehardware,andGo’sevaluationfunctioncouldchangedrasticallywithonestonechange.Recentworks[Maddi-sonetal.(2015);Clark&Storkey(2015)]showthatsearchisnotstrictlynec-essaryformachineGoplayers.A

3、purepattern-matchingapproach,basedonaDeepConvolutionalNeuralNetwork(DCNN)thatpredictsthenextmove,canperformaswellasMonteCarloTreeSearch(MCTS)-basedopensourceGoen-ginessuchasPachi[Baudis&Gailly(2012)]ifitssearchbudgetislimited.Weextendthisideainourbotnameddarkforest,whichreliesonaDCNNdesignedforlon

4、g-termpredictions.Darkforestsubstantiallyimprovesthewinrateforpattern-matchingapproachesagainstMCTS-basedapproaches,evenwithloosersearchbudgets.Againsthumanplayers,darkforestachievesastable1d-2dlevelonKGSGoServer,estimatedfromfreegamesagainsthumanplayers.Thissubstan-tiallyimprovestheestimatedranki

5、ngsreportedinClark&Storkey(2015),whereDCNN-basedbotsareestimatedat4k-5klevelbasedonperformanceagainstothermachineplayers.AddingMCTStodarkforestcreatesamuchstrongerplayer:withonly1000rollouts,darkforest+MCTSbeatspuredarkforest90%ofthetime;with5000rollouts,ourbestmodelplusMCTSbeatsPachiwith10,000rol

6、louts95:5%ofthetime.1INTRODUCTIONForalongtime,computerGoisconsideredtobeagrandchallengeinartificialintelligence.Fig.1showsasimpleillustrationofthegameofGo.Twoplayers,blackandwhite,placestonesatinter-sectionsinturnona19x19board(Fig.1(a)).Blackplaysfirstonanemptyboard.A4-connectedarXiv:1511.06410v1[cs

7、.LG]19Nov2015componentofthesamecoloriscalledagroup.Thelibertiesofagroupisthenumberofitsneigh-boringemptyintersections(Fig.1(b)).Agroupiscapturedifitslibertiesarezero.Thegoalofthegameistocontrolmoreterritorythanth

当前文档最多预览五页，下载文档查看全文

此文档下载收益归作者所有

当前文档最多预览五页，下载文档查看全文
温馨提示：
1. 部分包含数学公式或PPT动画的文件，查看预览时可能会显示错乱或异常，文件下载后无此问题，请放心下载。
2. 本文档由用户上传，版权归属用户，天天文库负责整理代发布。如果您对本文档版权有争议请及时联系客服。
3. 下载前请仔细阅读文档内容，确认文档内容符合您的需求后进行下载，若出现内容与标题不符可向本站投诉处理。
4. 下载文档时可能由于网络波动等原因无法下载或下载错误，付费完成后未能成功下载的用户请联系客服处理。