Hurtig modstandermodellering i simplificeret poker

Rapid Opponent Modeling in Simplified Poker

Jensen, Flemming

2. term

Computer Science, Master

2009

134

I poker er en spilteoretisk optimal (GTO) strategi stærk mod ukendte modstandere. Mod spillere, der ikke er optimale, kan modellering af modstanderen afsløre svagheder, der kan udnyttes. Denne afhandling undersøger, hvor gennemførlige to tilgange til modstander-modellering er i Leduc Hold'em, en simplificeret én-mod-én (heads-up) udgave af Fixed Limit Texas Hold'em. Fordi mange virkelige spil kun varer få hundrede hænder, fokuserer vi på kampe med højst 500 hænder. Ved eksplicit modstander-modellering estimerer vi modstanderens strategi ud fra observerede handlinger og beregner derefter en best-response strategi—en strategi designet til at vinde mest mod det estimat. Ved implicit modstander-modellering vælger vi i stedet, blandt et fast sæt kandidatstrategier, den der klarer sig bedst mod modstanderen, uden at bygge en eksplicit model. Med en testramme gennemførte vi en bred vifte af eksperimenter. I vores setup virker den eksplicitte tilgang uigennemførlig inden for så korte kampe, mens den implicitte tilgang virker gennemførlig. Vi afslutter med at skitsere en ny teknik til modstander-modellering, som vil blive implementeret og testet fremover.

In poker, a game-theoretic optimal (GTO) strategy is strong against unknown opponents. Against players who are not optimal, modeling the opponent can reveal weaknesses to exploit. This thesis tests the feasibility of two opponent-modeling approaches in Leduc Hold'em, a simplified one-on-one (heads-up) version of Fixed Limit Texas Hold'em. Because many real games last only a few hundred hands, we focus on matches of at most 500 hands. In explicit opponent modeling, we estimate the opponent’s strategy from observed actions and then compute a best-response strategy—one designed to win the most against that estimate. In implicit opponent modeling, we instead pick, from a fixed set of candidate strategies, the one that performs best against the opponent, without building an explicit model. Using a test framework, we ran a wide range of experiments. In our setting, explicit modeling appears infeasible within such short matches, while implicit modeling appears feasible. We conclude by outlining a new opponent-modeling technique to be implemented and tested in future work.

[This abstract was generated with the help of AI]

Poker ; LIMIDs ; Game theory ; Opponent modeling ; Poker ; LIMIDs ; Spilteori ; Modstandermodellering

Download
View record in AAU Student Projects