Abstract:
The state-of-the-art within Artificial Intelligence has directly benefited from research conducted within the computer poker domain. One such success has been the the advancement of bottom up equilibrium finding algorithms via computational game theory. On the other hand, alternative top down approaches, that attempt to generalise decisions observed within a collection of data, have not received as much attention. In this thesis we examine top down approaches that use Case-Based Reasoning in order to construct strategies within the domain of computer poker. Our analysis begins with the development of frameworks to produce static strategies that do not change during game play. We trace the evolution of our case-based architecture and evaluate the effect that modifications have on strategy performance. The end result of our experimentation is a coherent framework for producing strong case-based strategies based on the observation and generalisation of expert decisions. Next, we introduce three augmentation procedures that extend the initial frameworks in order to produce case-based strategies that are able to adapt to changing game conditions and exploit weaknesses of their opponents. Two of the augmentation procedures introduce different forms of opponent modelling into the case-based strategies produced. A further extension investigates the use of transfer learning in order to leverage information between separate poker sub-domains. For each poker domain investigated,we present results obtained fromthe Annual Computer Poker Competition, where the best poker agents in the world are challenged against each other. We also present results against a range of human opponents. The presented results indicate that the top down case-based strategies produced are competitive against both human opposition, as well as state-of-the-art, bottom up equilibrium finding algorithms. Furthermore, comparative evaluations between augmented and non-augmented frameworks show that strategies which have been augmented with either transfer learning or opponent modelling capabilities are typically able to outperformtheir non-augmented counterparts.