Fast Convergence in Semi-Anonymous Potential Games

Abstract

Log-linear learning has been extensively studied in both the game theoretic and distributed control literature. It is appealing for many applications because it often guarantees that the agents' collective behavior will converge in probability to the optimal system configuration. However, the worst case convergence time can be prohibitively long, i.e., exponential in the number of players. We formalize a modified log-linear learning algorithm whose worst case convergence time is roughly linear in the number of players. We prove this characterization for a class of potential games where agents' utility functions can be expressed as a function of aggregate behavior within a finite collection of populations. Finally, we show that the convergence time remains roughly linear in the number of players even when the players are permitted to enter and exit the game over time.