Multi-Agent Task Planning Algorithm Based on Region Parameter Sharing
Wen Zhang, Qing Weng, Kun Li, Xianchao Cao, Jing Wang · 2023
Aiming at the environmental nonstationarity problem caused by multiple agents interacting with the environment simultaneously, multi-agent task planning algorithms typically use a global information exchangeable model. However, this approach doesn’t consider the limitations of communication among agents in real-world applications, leading to poor algorithm practicality, while facing problems such as low training efficiency and slow convergence speed. For this, we first propose a multi-topology training with decentralized execution (MTTDE) mechanism that allows agents to dynamically choose between partial or centralized training methods based on their real-time communication conditions. Based on this, we introduce a new regional parameter sharing (RPS) method that simplifies the partial computation of network parameters in policy iteration. Finally, based on RPS method we propose an efficient multi-topology region parameter sharing multi-agent deep deterministic policy gradient (MTRPS-MADDPG) algorithm for environmental communication constraints. Experimental results on a multi-agent cooperative patrol task show that the proposed algorithm significantly improves the training efficiency and convergence speed compared with the traditional algorithm under the communication constraint and without affecting performance.