基于深度 Q 网络的多智能体逃逸算法设计

At present, the problem of multi-agent pursuit-evasion game is usually studied in the two-dimensional plane, and the movement of the evader is not constrained. At the same time, one problem is that it is difficult for traditional methods to design control strategy without accurate model. Therefore, this paper proposes a multi-agent evasion algorithm based on deep Q-network when the motion of evader is constrained in three-dimensional space. The proposed algorithm is a decentralized algorithm, and the evader obtains the desired evasive strategy by exploring and learning the environment. In order to improve the learning efficiency, the agent strategy learning is divided into two stages according to the difficulty of the task, and the corresponding reward function is designed to guide the agent to explore the desired evasive strategy. The simulation results show that the effect of the evasive strategy obtained by the algorithm is stable, and the algorithm has generalization ability, and the evader can successfully evade after changing certain initial position conditions.