On Passivity and Reinforcement Learning in Finite Games

Abstract

We use a passivity-based methodology for the analysis and design of reinforcement learning in multi-agent games. We consider an exponentially-discounted reinforcement learning scheme, and show that convergence can be guaranteed for the class of games characterized by the monotonicity property of their (negative) payoff. We further exploit passivity properties to propose a class of higher-order schemes that preserve convergence properties.