42
By David Anderson SZTAKI (Budapest, Hungary) WPI D2009

By David Anderson SZTAKI (Budapest, Hungary) WPI D2009

  • Upload
    others

  • View
    3

  • Download
    0

Embed Size (px)

Citation preview

ByDavidAndersonSZTAKI(Budapest,Hungary)WPID2009

  1997,DeepBluewonagainstKasparov  AverageworkstationcandefeatbestChessplayers

  ComputerChessnolonger“interesting”  Goismuchharderforcomputerstoplay

  Branchingfactoris~50‐200versus~35inChess  Positionalevaluationinaccurate,expensive  Gamecannotbescoreduntiltheend

  BeginnerscandefeatbestGoprograms

  Two‐player,totalinformation  Playerstaketurnsplacingblackandwhitestonesongrid

  Boardis19x19(13x13or9x9forbeginners)  Objectistosurroundemptyspaceasterritory  Piecescanbecaptured,butnotmoved Winnerdeterminedbymostpoints(territorypluscapturedpieces)

Imagefromhttp://ict.ewi.tudelft.nl/~gineke/

 Minimax/α‐βalgorithmsrequirehugetrees  Treedepthcannotbecuteasily

 Monte‐Carlonowmorepopular  Simulaterandomgamesfromthegametree  Useresultstopickbestmove

  Twoareasofoptimization  Discoveryofgoodpathsinthegametree  Intelligenceofrandomsimulations▪  Randomgamesareusuallybogus

  Needtobalancebetweenexploration…  Discoveringandsimulatingnewpaths

  Andexploitation…  Simulatingthemostoptimalpath

  BestmethodiscurrentlyUCTgivenbyLeventeKocsisandCsabaSzepesvári.

Sayyouhaveaslotmachinewithaprobabilityofgivingyoumoney.Youcaninferthisprobabilitythroughexperimentation.

Whatiftherearethreeslotmachines,andeachhasadifferentprobability?

Youneedtochoosebetweenexperimenting(exploration)andgettingthebestreward(exploitation).

UCBalgorithmbalancestheseproblemstominimizelossofreward.

UCTappliesUCBtogameslikeGo,decidingwhichmovetoexplorenextbytreatingitlikethebanditproblem.

  Startswithone‐leveltreeoflegalboardmoves

  PicksbestmoveaccordingtoUCBalgorithm

  RunsMonte‐Carlosimulation,updatenode’swin/loss.

  ThisisoneiterationoftheUCTprocess.

  Ifnodegetsvisitedenoughtimes,startlookingatitschildmoves

  UCTdivesdeeper,eachtimepickingthemost“interesting”move.

  Eventually,UCThasbuiltalargetreeofsimulationinformation

  UCTisnowinmostmajorcompetitiveprograms

  “MoGo”usedUCTtodefeataprofessional  Used800‐nodegridanda9stonehandicap

 Muchresearchnowfocusedonimprovingsimulationintelligence

  Policydecideswhichmovetoplaynextinarandomgamesimulation

  HighstochasticitymakesUCTlessaccurate  Takeslongertoconvergetocorrectmove

  ToomuchdeterminismmakesUCTlesseffective  DefeatspurposeofMonte‐Carlosearch Mightintroduceharmfulselectionbias

  CertainshapesinGoaregood  “Hane”hereisastrongattackonB

  Othersarequitebad!  B’s“emptytriangle”istoodenseandwasteful

 MoGousespatternknowledgewithUCT  Hand‐crafteddatabaseof3x3interestingpatterns  Doubledsimulationwin‐rateaccordingtoauthors

  Canpatternknowledgebetrainedautomaticallyviamachinelearning?

  Paper“Monte‐CarloSimulationBalancing”  (byDavidSilverandGeraldTesauro)  Policiesaccumulateerrorwitheachmove  Strongpoliciesminimizethiserror,butnotthewhole‐gameerror

  Proposesalgorithmsforminimizingwhole‐gameerrorwitheachmove

  Authorstestedon5x5Gousing2x2patterns  Foundthatbalancingwasmoreeffectiveoverrawstrength

  Implementedpattern‐learningalgorithmsin“Monte‐CarloSimulationBalancing”  Strength:Apprenticeship  Strength:PolicyGradientReinforcement  Balance:PolicyGradientSimulationBalancing  Balance:Two‐StepSimulationBalancing

  Used9x9Gowith3x3patterns

  Usedamateurdatabaseof9x9gamesfortraining

 Mention‐worthymetrics:  Simulationwinrateagainstpurelyrandom  UCTwinrateagainstUCTpurelyrandom  UCTwinrateagainstGNUGo

  Simplestalgorithm  Looksateverymoveofeverygameinthetrainingset  Highpreferenceforchosenmoves  Lowpreferenceforunchosenmoves

  Stronglyfavoredgoodpatterns  Over‐training;poorerrorcompensation

  Valuesconvergetoinfinity

0

10

20

30

40

50

60

70

80

Playout UCTvslibEGO UCTvsGNUGo

Winrate(%

)

GameType

ApprenticeshipvsPureRandom

PureRandom

Apprenticeship

  Playsrandomgamesfromthetrainingset  Ifthesimulationmatchestheoriginalgameresult,patternsgethigherpreference

  Otherwise,lowerpreference  Resultswerepromising

0

10

20

30

40

50

60

70

Playout UCTvslibEGO UCTvsGNUGo

Winrate(%

)

GameType

ReinforcementvsPureRandom

PureRandom

Reinforcement

  Foreachtraininggame…  Playsrandomgamestoestimatewinrate  Playsmorerandomgamestodeterminewhichpatternswinandlose

  Givespreferencestopatternsbasedonerrorbetweenactualgameresultandobservedwinrate

  Usually,stronglocalmoves  Seemedtolearngoodpatterndistribution  Aggressivelyplayeduselessmoveshopingforanopponentmistake

  Poorconsiderationofthewholeboard

0

10

20

30

40

50

60

Playout UCTvslibEGO UCTvsGNUGo

Winrate(%

)

GameType

SimulationBalancingversusPureRandom

PureRandom

SimulationBalancing

  Picksrandomgamestates  Computesscoreestimateofeverymoveat2‐plydepth

  Updatespatternpreferencesbasedontheseresults,usingactualgameresulttocompensateforerror

  Gamescoreishardtoestimate,usuallyinaccurate

  Extremelyexpensive;10‐30sectoestimatescore

  Gamescoredoesn’tchangemeaningfullyformanymoves

  Probablydoesnotscaleasboardsizegrows

0

10

20

30

40

50

60

70

Playout UCTvslibEGO UCTvsGNUGo

Winrate(%

)

GameType

TwoStepBalancingvsPureRandom

PureRandom

TwoStepBalancing

0

10

20

30

40

50

60

70

80

Playout UCTvslibEGO UCTvsGNUGo

Winrate(%

)

GameType

AlgorithmResults

PureRandom

Apprenticeship

Reinforcement

SimulationBalancing

TwoStepBalancing

  Reinforcementstrongest  Allalgorithmscapableofverydeterministicpolicies

  HigherplayoutwinratesweretoodeterministicandthususuallybadwithUCT

  Gomaybetoocomplexforthesealgorithms  Optimizingself‐playdoesn’tguaranteegoodmoves

  LeventeKocsis

  SZTAKI

  ProfessorsSárközyandSelkow

  Algorithmgenerateslistofpatterns  Eachpatternhasaweight/value  Policylooksatopenpositionsontheboard  Getsthepatternateachopenposition  Usesweightsasaprobabilitydistribution