Evolved neural networks learning Othello strategies

Abstract

Evolutionary computation was used to train neural networks to learn the play the game of Othello. Each neural network represents a strategy based on board evaluations of the game tree generated by a minimax search algorithm. Networks competed against each other in tournament play and selection used to eliminate those that performed poorly relative to other networks. Self-adaptation was used to mutate the weights and biases of surviving neural networks to generate offspring. By monitoring the evolutionary behavior over 1000 generations through game competitions with computer players playing at higher ply-depths using deterministic evaluations, the networks are shown to coevolve with the style of game play progressing from random to positional and finally to mobility strategy.