Evolutionary Methods for Learning No-limit Texas Hold’em Poker

No-Limit Texas Hold’em is a stochastic game of imperfect information. Cards are dealt randomly, and players try to hide which cards they are holding from their opponents. Randomness and imperfect information give Poker, in general, and No-Limit Texas Hold’em, in particular, a very large decision space when it comes to making betting decisions. Evolutionary algorithms and artificial neural networks have been shown to be able to find solutions in large and non-linear decision spaces, respectively. A hybrid method known as evolving neural networks is used to allow No-Limit Texas Hold’em Poker playing agents to make betting decisions. The evolutionary heuristics of halls of fame and co-evolution extend the evolving neural networks to counter evolutionary forgetting and improve the quality of the evolved agents. The appropriateness of a tournament-based fitness function for the evolutionary algorithms is investigated. The results show that the tournament-based fitness function is the most appropriate of the tested fitness functions. Furthermore, the results show that the use of co-evolution and a hall of fame increases the quality of the evolved agents against a number of benchmark agents. The use of co-evolution and a hall of fame separately show that the hall of fame has the greater influence on the evolution of the agents, and the addition of co-evolution to the hall of fame has little added benefit.