Expectation-based learning in two-person finite games

Abstract

Learning pure-strategy equilibrium of normal form bi-matrix games in the assumption of knowledge of own-payoffs and no knowledge of rival strategies is considered. An original learning algorithm based on mixed best-reply to expectations is proposed. Global convergence is ensured for a new class of games including but not restricted to potential games. Results of classic Linear-Reward Inaction schemes are significantly improved at the modest cost of knowledge of own payoffs.