Modeling opponent learning in multiagent repeated games

Abstract Multiagent reinforcement learning (MARL) has been used extensively in the game environment. One of main challenges MARL is that environment agent system dynamic, and other agents are also updating their strategies. Therefore, modeling opponents’ process adopting specific strategies to shape an effective way obtain better training results. Previous studies such as DRON, LOLA SOS approximated opponent’s gave applications. However, these modeled only transient changes opponent lacked stability improvement equilibrium efficiency. In this article, we design MOL (modeling learning) method based on Stackelberg game. We use best response theory approximate preferences for different actions explore stable with higher rewards. find achieves results several games classical structures (the Prisoner’s Dilemma, Leader Stag Hunt 3 players), randomly generated bimatrix games. performs well competitive played against opponents converges points score above Nash repeated environments. The may provide a reference definition multiagent systems, contribute objectives avoid local disadvantageous improve general