基于Markov对策和强化学习的多智能体协作研究

MAS的协作机制研究，当前比较适用的研究框架是非零和Markov对策及基于Q－算法的强化学习。但实际上在这种框架下的Agent强调独立学习而不考虑其他Agent的行为，故MAS缺乏协作机制。并且，Q－算法要求Agent与环境的交互时具有完备的观察信息，这种情况过于理想化。文中针对以上两个不足，提出了在联合行动和不完备信息下的协调学习。理论分析和仿真实验表明，协调学习算法具有收敛性。

Research on Multiagent Cooperation with Markov Game and Reinforcement Learning

Non zero sum Markov game and reinforcement learning based on Q algorithm is a feasible frame for the research on the mechanism of multiagent system's cooperation. In fact, the independent learning is accentuated for agent regardless of other agents' actions under this frame. So, the mechanism of cooperation is deficient. And, it is over idealized that the perfect observed information is required when agents are interacting with environment. In the paper, cooperated learning under joined action and imperfect information was proposed for solving these two problems. Convergence of the improving algorithm was proved.