This paper provides a number of contributions to improving the play of a poker-playing program. Not in the least is the effective use of decision networks in improving the performance of BPP significantly.
The use of the decision networks allowed for the removal and refinement of a number of expert-knowledge components evident within the previous betting strategy based upon pot odds and betting curves. The added accuracy provided in determining the expected winnings for each possible action caused it to out-perform the previous betting technique convincingly. Decision networks provided a more intuitive means for modeling how actions are selected, with the technique proposed having the advantage of easily being able to be used in a mixed or deterministic strategy depending on the specifications of the domain, and which strategy produces the better results. The use of decision networks in selecting a betting action has been shown experimentally to significantly improve the quality of betting decisions made by the Bayesian Poker Player and has proved that decision networks are an appropriate and effective way to represent games which require decisions to be made under conditions of uncertainty and imperfect information achieving high performance with minimal expert knowledge.
To master the game of poker, one must be adaptive. Any form of deterministic play can and will be exploited by a good opponent. A player must change their style based on the dynamic game conditions observed over a series of hands. This work has made some progress towards achieving a poker-playing program that can learn and adapt. BPP successfully uses opponent modeling to improve its play. However, it is abundantly clear that these are only the first steps, and there is considerable room for improvement. Poker is a complex game. Strong play requires the player to excel in many different aspects of the game. Developing BPP is an on going process. One aspect of the program is improved until it becomes apparent that another aspect is the performance bottleneck. That problem is then tackled until it is no longer the limiting factor, and new weaknesses in the program's play are revealed. This attempt at improving the opponent modeling aspect of the game has shown some improvements, but it is evident that it is holding back performance and should be of focus for future development. Not in the least, determining a less biased and more descriptive means of representing how an opponent's actions are modeled within the game.
As experimental results point out, the improved version of BPP wins more (plays better) than the original version. However, the program is still a long way from achieving the goal of world class poker play, but many significant improvements have being made to its performance and this paper has presented a number of effective ways by which play was significantly improved.