Multi-Agent Path Finding in Dynamic Environment Guided by Historical Experience

Multi-agent path finding (MAPF) is a challenging problem in large-scale dynamic environments. automated guided vehicles (AGVs) need to reach given targets within a limited number of steps while avoiding potential conflicts with other AGVs or static shelves. For MAPF, DRL methods typically combine local observations and the global map, allowing AGVs to obtain obstacle information in the map, the next actions of other AGVs and their respective targets during operation. However, this approach requires timely access to a large amount of global information, especially as the map size increases. AGVs processors need to perform a large amount of computation in a short time, making it difficult for AGVs to respond. To address this issue, we propose a history-informed intelligent navigation with deep reinforcement learning method (HIRL). This method uses the historical exploration experience of AGVs, combines the historical experience path and local observation in the action, and jointly determines the current path action. We apply HIRL to solve the MAPF problem in a fully distributed manner and evaluate our method on different map types, obstacle densities, and AGVs numbers. Experimental results show that HIRL exhibits good generalization and better performance under the increasing number of AGVs, which is better than the existing distributed methods.