【摘要】 提出了一种综合了模块化结构、利益分配学习以及对手建模技术的多agent强化学习方法,利用模块化学习结构来克服状态空间的维数灾问题,将Q-学习与利益分配学习相结合以加快学习速度,采用基于观察的对手建模来预测其他agent的动作分布.追捕问题的仿真结果验证了所提方法的有效性.更多还原

【Abstract】 A novel multi-agent reinforcement learning approach is proposed to learn the coordinated behaviors among cooperative agents team.The proposed approach combines advantages of the modular architecture,profit-sharing learning and opponent modeling technique in a single multi-agent framework.Simulation results on the pursuit problem show that the proposed learning approach has faster convergence speed and more optimal policy over conventional modular Q-learning algorithms.更多还原