GA-Gammon: A Backgammon Player Program Based on Evolutionary Algorithms

In this paper we describe a genetic algorithm approach able to confection strong backgammon automata players. We first prepared an initial vector of weights representing a set of heuristic strategies suggested by expert human players. Then, employing a genetic algorithm approach we were able to fine tune such initial vector of weights by repeatedly testing it against Pubeval, a strong benchmark player program. The vector of weights was therefore used as an evaluation function for performing a genetic heuristic selection of the best board positions during a game. Best GA-Gammon individuals so obtained were tested in separated 5000-game tournaments against Pubeval itself, and Fuzzeval, a fuzzy controllerbased player. Our experimental results indicate that the best individuals generated by GAGammon show similar performance than Pubeval. Furthermore, GA-Gammon consistently outperforms Fuzzeval.