关键词: 多机器人路径规划, 深度强化学习, 注意力机制, 通信

Abstract: Multi-agent path planning in a dynamic environment has great practical value and academic value in the field of multi-agent. In order to reduce the blocking rate of multi-agent path planning in a dynamic environment, under the framework of the deep reinforcement learning algorithm Actor-Critic, In this paper, a distributed multi-agent path planning method based on request and response communication mechanism and local attention mechanism is designed, named Distributed Communication and local Attention Based Multi-agent Path Finding(DCAMAPF). In the Actor network, based on the request and response communication mechanism, each agent requests the local observation information and action information of other agents in the field of view, and then plans a coordinated action policy. In the Critic network, each agent dynamically assigns attention weights to the local observation and action information of other agents that successfully respond within the field of view based on the local attention mechanism. Compared with the traditional dynamic path planning method D* Lite, the latest distributed reinforcement learning method Mapper, and the latest centralized reinforcement learning method AB-Mapper, in the discrete initialization environment of DCAMAPF, the difference in blocking rate was reduced by about 6.91%, 4.97%, and 3.56%. In the centralized initialization environment, blocking can be avoided more efficiently, and the difference in blocking rate is reduced by about 15.86%, 11.71%, and 5.54%, and less computing cache is occupied.

Key words: Keywords: Multi-agent path planning, deep reinforcement learning, attention mechanism, communication