Cooperative pursuit-strategy based on RL for pursuer-evader games

This paper presents a learning methodology for multiple pursuers to collaborate in chasing an evader within environments characterized by dynamic obstacles. Leveraging a variational autoencoder framework for the obstacle detection, the multi-agent twin delayed deep deterministic policy gradient algorithm is utilized to train the pursuers, while the deep deterministic policy gradient algorithm is employed to direct the evader, forming a comprehensive learning-based pursuit-evasion model. A decoupled reward structure is implemented to balance cooperative and individual incentives, thereby enhancing learning effectiveness.