Evolving heuristic based game playing strategies for checkers incorporating reinforcement learning

Abstract

The research presented in this paper forms part of a larger initiative aimed at creating a general game player for two player zero sum board games. In previous work, we have presented a novel heuristic based genetic programming approach for evolving game playing for the board game Othello. This study extends this work by firstly evaluating it on a different board game, namely, checkers. Secondly, the study investigates incorporating reinforcement learning to further improve evolved game playing strategies. Genetic programming evolves game playing strategies composed of heuristics, which are used to decide which move to make next. Each strategy represents a player. A separate genetic programming run is performed for each move of the game. Reinforcement learning is applied to the population at the end of a run to further improve the evolved strategies. The evolved players were found to outperform random players at checkers. Furthermore, players induced combining genetic programming and reinforcement learning outperformed the genetic programming players. Future research will look at further application of this approach to similar non-trivial board games such as chess.