Research Seminar - September 17, 1999

Abstract

Poker is more than a game of chance. Good play requires not only a degree of bluffing, but the ability to deduce opponent playing styles in order to exploit weaknesses. This requires adaptive learning.

Evolutionary algorithms model the process of natural selection in computers. Utilising the implicit learning characteristics of evolution, evolutionary algorithms model adaptation in dynamic environments.

I shall describe a new learning model using evolutionary algorithms suitable for designing adaptive computer poker players. I will identify several important poker principles and use these as the basis for a hypercube of evolving populations of poker competences. I will discuss experiments that demonstrate the emergent adaptive behaviour of the evolving computer poker players, and in particular, show that the evolving poker players develop different techniques to maximise winnings against different opponents. A comparison of the strategies evolved by the computer poker players with a competent fixed strategy will demonstrate the improved performance of this approach.