The role of behavioral diversity and difficulty of opponents in coevolving game-playing agents

Abstract

Generalization performance of learning agents depends on the training experience to which they have been exposed. In gameplaying domains, that experience is determined by the opponents faced during learning. This analytical study investigates two characteristics of opponents in competitive coevolutionary learning: behavioral diversity and difficulty (performance against other players). To assess diversity, we propose a generic intra-game behavioral distance measure, that could be adopted to other sequential decision problems. We monitor both characteristics in two-population coevolutionary learning of Othello strategies, attempting to explain their relationship with the generalization performance achieved by the evolved solutions. The main observation is the existence of a non-obvious trade-off between difficulty and diversity, with the latter being essential for obtaining high generalization performance.