EvoMCTS: A Scalable Approach for General Game Learning

In this paper, we present the application of genetic programming as a generic game learning approach to zero-sum, deterministic, full-knowledge board games by evolving board-state evaluation functions to be used in conjunction with Monte Carlo tree search (MCTS). Our method involves evolving board-evaluation functions that are then used to guide the MCTS playout strategy. We examine several variants of Reversi, Dodgem, and Hex using strongly typed genetic programming, explicitly defined introns, and a selective directional crossover method. Our results show a proficiency that surpasses that of baseline handcrafted players using equal and in some cases greater amounts of search, with little domain knowledge and no expert domain knowledge. Moreover, our results exhibit scalability.