Distributing Rewards by Strategic Knowledge based on Nash-Q Learning

摘要

In this investigation, we examine collaboration approach to reward distribution in repeated general-sum stochastic games by multiple game players in terms of position and rewards. There have been several investigation of reward distribution discussed so far, and reinforcement has been considered useful since no knowledge is needed in advanced and better decision can be extracted while learning. Among others, Q-learning has been paid much attention under single agent environment. However, under multi-agent environment, we don't have sharp targets to this problem, what is the most optimal principle? In this work, we discuss how to distribute reward thoroughly by considering as general stochastic games based on theory of games. That is, we introduce Nash-Q approach which combines Nash equilibrium with Q-learning. We show the new approach provides us with new strategic solution. We discuss some experiments of rather complicated games (game of life) to see the usefulness of the approach.