A multi-agent coordination framework based on Markov games

Based on the analysis of the reinforcement learning and Markov games, the paper proposes a layered multi-agent coordination framework. Based on the multi-agent´s interaction of competition and cooperation, this coordination framework adopts the zero-sum Markov game in higher layer to compete with opponent and adopts the team Markov game in lower layer to accomplish the team´s cooperation. This coordination framework is applied to Robot Soccer. The results of the experiment illuminate that our proposed method is better than the traditional multi-agent learning.