Search

A multi-agent reinforcement learning method based on action prediction of other agents is proposed. In the multi-agent system, the action selection of a learning agent must be effected by other agents' action. Therefore, joint-state and joint-action are involved in reinforcement learning system. The method of agent action prediction based on the probabilistic neural network is proposed. So the joint-action is formed and multi-agent reinforcement learning is implemented. Furthermore, the application of the proposed method in cooperation strategy learning of robot soccer is studied. Through mutually learning with environment, multiple robot system can master behavior strategy and realize multiple robots coordination and cooperation. [ABSTRACT FROM AUTHOR]

Considering the aviation cost and combat effect of aerocrafts team, this thesis investigated the way of aerocrafts' cooperative path planning based on task assignment, constructed aerocrafts team's cooperative path planning model and designed decomposition algorithm. This can effectively balance multiple objective presence, get reasonable task allocation and path planning schemes. The simulation results show that this method can not only guarantee the cooperative paths of aircrafts reasonable, but also get the best combat effect and effectively improve the combat benefit-cost ratio. [ABSTRACT FROM AUTHOR]