S.-K. Chai, J.J. Salerno, and P.L. Mabry (Eds.): SBP 2010, LNCS 6007, pp. 398–405, 2010.
© Springer-Verlag Berlin Heidelberg 2010
Opponent Classification in Poker
Muhammad Aurangzeb Ahmad and Mohamed Elidrisi
Department of Computer Science and Engineering, University of Minnesota
{mahmad,elidrisi}@cs.umn.edu
Abstract. Modeling games has a long history in the Artificial Intelligence com-
munity. Most of the games that have been considered solved in AI are perfect
information games. Imperfect information games like Poker and Bridge repre-
sent a domain where there is a great deal of uncertainty involved and additional
challenges with respect to modeling the behavior of the opponent etc. Tech-
niques developed for playing imperfect games also have many real world appli-
cations like repeated online auctions, human computer interaction, opponent
modeling for military applications etc. In this paper we explore different tech-
niques for playing poker, the core of these techniques is opponent modeling via
classifying the behavior of opponent according to classes provided by domain
experts. We utilize windows of full observation in the game to classify the
opponent. In Poker, the behavior of an opponent is classified into four standard
poker-playing styles based on a subjective function.
Keywords: Opponent Classification, Opponent Modeling, Poker.
1 Introduction
A game is a structured or a semi-structured interaction between two or more entities
where there usually some incentives involved in playing the game. Games can be
classified into games with perfect information and games with imperfect information.
In games with perfect knowledge the complete state of the game is observable. Ex-
amples of such games include Tic-Tac-Toe, Backgammon, Go, Chess etc. The most
common approach in solving these games is to build a game tree and determine the
best strategy to use given one’s position in the game tree. Such games are readily
amenable to brute force search approaches. A history of the study of this class of
games demonstrates that many of these games were solved as the computational
resources became available.
Games with imperfect information on the other hand are fundamentally different
because the complete state of the game is not observable. In many such games either
the search space is intractable or even in cases of tractability many of the techniques
employed in games with perfect information do not do well in games with imperfect
information. In theory even if one could state the optimal strategy for imperfect
information games like Poker, computing the strategy would be prohibitively expen-
sive and given the large uncertainties involved may not always work. Thus alternative
techniques to traditional approaches used for perfect information games have to be
employed. Poker is a game with imperfect information often used as a test bed for