Soft Actor-Critic Based Multi-drones Pursuit-Evasion Differential Game with Obstacles

The multi-drones pursuit-evasion differential game problem (PEDG) has been one of the important problems of multi-drones system. In this paper, we study the multi-drones pursuit-evasion differential game in an environment with obstacles. In order to solve the partially observable problem in the multi-agent game, this paper adopts the Soft Actor-Critic, a deep reinforcement learning algorithm and combines centralized learning and distributive execution. In order to improve the generalization of the trained model in the obstacle environment, the obstacles are randomly generated during training, and the curriculum learning method is used during training by gradually increasing the number of obstacles in order to accelerate convergence. In this paper, the trained model is analyzed for generalization and tested in various environments with randomly generated obstacles, and all of them achieve good performance. This paper compares the trained model with the evader with random strategy, which shows the existence of the game to some extent.