Modifications to BPP

BPP was a intermediate level poker player as shown by the experimental results in [Korb, Nicholson, and JitnahKorb et al.1999]. Also, BPP demonstrated the benefits of using opponent modeling and a randomised betting strategy. However, a number of weaknesses were identified that were hampering the overall performance of the program. These issues led to a redesign of BPP's architecture and the facilitation of new components and techniques to the program.

The underlying network structure required improvement as the original design had several limitations. First, being coded entirely in LISP, the Bayesian networks used by BPP were hard to maintain and improve. This effectively restricted the network structure to a polytree form, unable to represent the perceived dependence of opposing players hands. It was also noted that opponent modeling implemented by BPP was still very weak, not distinguishing and exploiting the behaviours of individual opponents. The course approximations and expect knowledge employed by in BPP's betting strategy also hindered performance tending to return a sub-optimal view of its expected winnings.

The improved version of BPP described in this thesis makes a number of fundamental changes to the original architecture. First, it introduces an improved belief network structure to the program. Chapter 4 explains the modifications and improvements made in detail. Second, the ways in which opponent actions are modeled and a specific opponent modeling scheme are presented (see Chapter 5). Finally, a totally new betting strategy using decision networks is explained and the relative merits are discussed in Chapter 6.