Learning in Multiplayer Stochastic Games
Howard M. Schwartz · 2014
In early work on multiagent reinforcement learning (MARL) for stochastic games, it was recognized that no agent works in a vacuum. This chapter reviews some existing reinforcement learning algorithms in stochastic games. It analyzes these algorithms based on their applicability, rationality, and convergence properties. The chapter introduces a grid version of Isaacs' game called the grid game of guarding a territory. It is a two-player zero-sum stochastic game where the defender plays against the invader in a grid world. The chapter then explains how the players learn to play the game using MARL algorithms. It applies two reinforcement learning algorithms to this game and tests the performance of these learning algorithms based on the convergence and rationality properties.