Multi-Agent Path Finding Based on Graph Neural Network

In order to improve the efficiency of logistics warehouses, multi-agent path finding (MAPF) has become a hot research. Although most of the current decentralized MAPF algorithms can obtain good results in experiments, the agents in the system do not communicate with each other, and do not fully utilize the scalability of decentralized control. The decentralized framework of this paper combines convolutional neural networks (CNN) and improved graph neural networks (GNN) to train a decentralized controller to simulate a coupled central controller. The decentralized framework based on GNN is used to find the path of the multi-agent system, and the optimal or near-optimal agent action sequence is obtained. This paper assumes that under the premise that the agent's hardware conditions are limited and can only perceive the local environment within the field of view. And realize efficient collision-free path planning by exchanging local information observed by agents through the GNN. Experiments show that the results obtained by the path finding algorithm based on GNN in this study are close to the optimal results, and have good scalability.