Learning Cooperative Strategies in Multi-Agent Encirclement Games with Faster Prey Using Prior Knowledge

Multi-agent encirclement with collision avoidance constitutes a common challenge in the multi-agent confrontation domain, wherein the focus lies in the development of cooperative strategies among agents. Previous studies encountered difficulties in addressing the dynamic encirclement of faster prey in obstacles environment. This paper introduces a novel multi-agent deep reinforcement learning approach based on prior knowledge. It is dedicated to exploring the multi-agent encirclement with collision avoidance task involving slower multiple pursuers collaboratively encircling faster prey in an obstacles environment. Firstly, the utilization of the classic Apollonius circle theory as prior knowledge guides agent action selection, narrows the exploratory action space, and accelerates the learning of strategies. Subsequently, the variance descriptor restricts the motion direction of pursuers, thus ensuring that pursuers continuously narrow the encirclement until the prey is successfully encircled. Finally, experiments in an obstacles environment were conducted to validate the proposed method. The results indicate that our method can acquire an effective encirclement strategy, with an encirclement success rate exceeding that of previous methods by more than 10