Abstract:

This paper aims at the learning process of a kind of cooperative teams, which pursue the maximum  benefit of a whole system. We propose a new cooperat ive reinforcement learning method based on the stochastic game in multi-agent systems. Each agent of the team decides its behaviors after forecasting the behavior strategy of acquaintances according to the stochastic game structure and their historical behaviors, and then a jointly optimal behavior stra tegy is obtained.

Key words: reinforcement learning, multi-agent system, stochastic game, cooperation