Learning a Reversi Board Evaluator with Minimax

A board position evaluator is a crucial component for strong computer play in many games such as checkers, chess, and Reversi. The board evaluator is typically trained using pre-existing game data, an approach which is generally non-optimal, especially in the early stages of a game. Instead, we propose a new method which relies on Minimax search to train a series of models backwards from the endgame, propagating information about endgame scoring backwards to earlier positions. In the limit of perfect models, our method is optimal, converging to the full Minimax board evaluation. We test our method experimentally by training a simple Reversi model using both our Minimax method and high-level tournament game data. When played against each other, our model outperforms the game data model, averaging 5 more stones per game, and winning approximately 60% of the matches.