Emergency Path Planning Research under Disaster Conditions Based on Reinforcement Learning and Layering Partitioning-Bidirectional A* Algorithm

Under disaster emergency conditions, fast and reliable vehicle emergency routes are a crucial foundation for conducting emergency dispatch and rescue operations. This paper introduces a global emergency path planning method based on reinforcement learning, which describes the mutual feedback process between road network traffic status, disaster relief demands, and path planning. We used a Transformer model based on self-attention mechanisms and trained the model using the Rollout baseline. The trained model is then utilized for global emergency path planning. Furthermore, the bidirectional A* algorithm is combined with the Layering Partitioning algorithm based on the road network topology model to optimize local paths. The effectiveness and feasibility of the algorithm are validated through a demonstration area.