Abstract:

A pursuit-evasion framework is proposed based on the adversarial evolutionary reinforcement learning algorithm for the problem of blue target intrusion in unmanned surface vehicle response to maritime emergencies. In order to improve the pursuit effect and generalization performance, the reinforcement learning method is used to increase the diversity of strategies for both the red unmanned surface vehicle and blue escape target, and the performance of the pursuit team is improved through the iterative adversarial evolution of both sides. For the pursuit team, considering that the individual may be damaged or exhausted of oil in the process of task execution, the multi-agent posthumous credit assignment algorithm is adopted and the residual-connected hidden long short-term memory network is introduced to improve the strategy network, and the obstacles such as islands and reefs are used to assist in improving the efficiency of unmanned surface vehicle encirclement and capture. Simulation results show that the adversarial evolution iterative training framework can effectively achieve the common progress of both pursuers and evaders, and the stability and convergence effect of the improved reinforcement learning algorithm are relatively strong. The proposed method demonstrates better intelligence and flexibility in addressing the problem of unmanned surface vehicle pursuit, and pursuit effect is significantly improved.

Key words: unmanned surface vehicle (USV), pursuit-escape, adversarial evolution, reinforcement learning