DHAA: Distributed heuristic action aware multi-agent path finding in high density scene

Multi-agent path finding (MAPF) in highly structured environments is an exciting and complex problem. Compared with lower-density environments, the problems of agent credit allocation and differentiated communication need to be effectively solved in a grid world with complex structures. Through assistance and decentralized training between agents, which local agent information is shared through communication. Therefore, this paper proposes a multi-agent path finding model for distributed heuristic behavior awareness in high-density scenes(DHAA). First, this paper uses a query mechanism to assign different weights to the agents in the neighborhood based on priority order to distinguish the importance of different information to the agents. Then, according to the differentiated communication and collision detection strategies, selective communication is carried out with agents that can change their strategies to achieve conflict avoidance. Finally, we encoder the values of the observation and state networks and input them into the deep Q network to get the following action. By embedding collision-free heuristic search paths into graph networks, which the model ability to handle long tasks is enhanced. Experiments show that the performance of the DHAA model proposed in this paper is close to the performance of the coupled centralized algorithm. We demonstrate the effectiveness of the DHAA model in high-density scenarios by exploring different benchmarks and ablation experiments. Experiments show that deep Q networks can reduce the average step size and differentiated communication can reduce bandwidth. The DHAA model also shows robustness in untrained maps, improving 23 % over the baseline model.