Emergent Correlated Equilibrium Through Synchronized Exploration
Published in RSS 2020 Workshop on Emergent Behavior in Human-Robot Systems, 2020
Recommended citation: M. Beliaev*, W. Wang*, D. Lazar, E. Biyik, D. Sadigh, R. Pedarsani. Emergent Correlated Equilibrium Through Synchronized Exploration. RSS 2020 Workshop on Emergent Behavior in Human-Robot Systems, July 2020.
Correlated equilibria strategies can be more prosocial, as they can achieve a larger expected sum of rewards compared to pure-strategy Nash equilibria. However, it can be difficult to reach correlated equilibria in multi-agent environments due to non-stationarity. We propose Synchronized \(\epsilon\)-Greedy Exploration, which builds on the commonly-used \(\epsilon\)-greedy exploration, and therefore can be generalized to stochastic games and used in any off-policy learning algorithm.