Improved Opponent Modeling in Poker

The game of poker has many properties that make it an interesting topic for arti cial intelligence (AI). It is a game of imperfect information, which relates to one of the most fundamental problems in computer science: how to handle knowledge that may be erroneous or incomplete. Poker is also one of the few games to be studied where deriving an accurate understanding of each opponent's style is an essential element to success. In developing a strong poker program, the opponent modeling method has always been a central component of the system. As other aspects of the program were improved, the techniques for modeling once again became a limiting factor to the overall level of play. As a result, the topic has been revisited. This paper reports on recent progress achieved by improved statistical methods, which were suggested by experiments using arti cial neural networks.