Cooperative control for multi-player pursuit-evasion games with reinforcement learning

In this paper, we consider a pursuit-evasion game in which multiple pursuers attempt to capture one superior evader. A distributed cooperative pursuit strategy with communication is developed based on reinforcement learning. The centralized critic and distributed actor structure and the learning-based communication mechanism are adopted to solve the cooperative pursuit control problem. Instead of using broadcast to share information among the pursuers, we construct the ring topology network and the leader-follower line topology network for communication, which could significantly reduce the complexity and save the communication and computation resources. The training algorithms for these two network topologies are developed based on the deep deterministic policy gradient algorithm. Furthermore, the proposed approach is implemented in a simulation environment. The training and evaluation results demonstrate that the pursuit team could learn highly efficient cooperative control and communication policies. The pursuers can capture a superior evader driven by an intelligent escape policy with a high success rate.