EvolutionofanEfficient SearchAlgorithmforthe - - PowerPoint PPT Presentation

• EvolutionofanEfficient

SearchAlgorithmforthe MateInNProbleminChess

AmiHauptmanandMosheSipper

BenGurionUniversity,Israel

2007“HUMIES” AWARDSFORHUMANCOMPETITIVERESULTS Monday,July9,2007

• …

GameStrategy= Search+Knowledge Search: Numberofnodesdeveloped Numberofnodesdeveloped Knowledge: Evaluationofnodes Evaluationofnodes Tradeoffbetweenthetwo

GamePlayingAI

• Powerfulcontemporaryengines
• Crafty

Crafty,Fritz,DeepJunior,… Lotsofsearch Lessknowledge

Intelligent?Hmmm…

Verylittlegeneralization Gobblescomputationalpower Deemedtheoreticallyuninteresting[Chomsky,93]

Chess:MachinePlayers

• Chess:HumanPlayers

Useproblemsolvingcognition Deeplyknowledge knowledge based based play Massiveuseofpatternrecognition patternrecognition; parallelism Alsousesearchbut

• Substantially

Substantially lessnodes(typicallydozens)

• Selective

Selective (only“good”)

• Moreefficient

Moreefficient: lessnodesfor“same” result

Goodsourceofinspirationforalgorithms

• OurGoal

Concentratingonendgameswepreviously:

evolvednodeevaluationfunction(knowledge)withGP Results:draworwinagainstCRAFTY,aworldclass chessengine Partofworkthatwona2005humiesmedal

Thiswork:Evolvethesearchalgorithmitself Evolvethesearchalgorithmitself Evolvebothsearchandknowledge bothsearchandknowledge,letting evolution evolution balancethetwo

• IncentiveforCurrentWork

Previouslyevolvedplayers:

Sometimesmiss(easy)shallowmates Scalingproblem:addingpiecestoboarddecreased scores

Evolvedplayersshouldrelymoreonsearch

Fullpureknowledgeplayerstillunattainable Searchmakesthestrongestengines

Problem:

Simplyadding search:tooslow(eachnodethoroughly examined)

• SOLUTION:

Balancingsearch&knowledgethroughevolution

evolution

• ProblemDomain

MateinN:Isthereaforcedwinsequencein maximum2*(N1)plies? Crucialtochessengines,searchedfarmore thoroughly CRAFTY:FordifficultN=5casessearchesover 106 nodes

• MajorResult

Evolvedsearchalgorithm: Numberofnodesdevelopedreducedby47% withrespecttoworld withrespecttoworld classengine(notsimple classengine(notsimpleαβ αβ) )

• ResultisHumanCompetitive

(H)resultholdsitsownorwinsaregulated competitioninvolvinghumanwrittencomputer programs (B)betterthanresultacceptedasanewscientific resultatthetime (D)resultispublishableinitsownright (F)betterthanresultconsideredanachievement atthetime (G)resultsolvesaproblemofindisputable difficultyinitsfield

• WhyisResultBest?

Difficultformosthumanchessplayers:

Musttrainintensivelynottomiss(and losegame)

Ourevolvedstrategiesimproveupononeoftop chessenginesinexistence(Crafty),representing manyhumanyearsofprogramming We’rebeatingthistopnotchengineinitsown “territory”:massivesearch Problemiscrucialtochessengines,therefore muchcomputationalpowerisexpended(e.g.,in suchpositions,DeepBlueexaminestwicethe normalnumberofnodes)

• WhyisResultBest?(cont’d)

Evolvingadynamicalgorithm(i.e.,aprocess) usuallyharderthanevolvingastaticstructure Wetookevolutiontothenextlevel:balancing searchandknowledge SurpassespreviousECsolutions

1. 1.Hardprobleminharddomainforman&machine(chess) Hardprobleminharddomainforman&machine(chess) 2. 2.Evolvedalgorithmbetter Evolvedalgorithmbetterthan than (most)humans (most)humans 3. 3.Evolvedalgorithm Evolvedalgorithm betterthanhuman betterthanhuman writtentopengine writtentopengine 4. 4.Evolutiontakentonextlevel Evolutiontakentonextlevel Inanutshell: Inanutshell:

• A.HauptmanandM.Sipper

Evolutionofanefficientsearchalgorithmforthemate innprobleminchess ProceedingsEuroGP2007,pp.7889, April2007 M.Sipper,Y.Azaria,A.Hauptman, &Y.Shichel Designinganevolutionarystrategizing machineforgameplayingandbeyond IEEETransactionsonSystems,Man, andCybernetics,PartC,37(4), July2007