Research on Emergency Rescue Path Optimization Method Based on Deep Reinforcement Learning

This paper presents the first attempt at an emergency rescue path optimisation method using deep reinforcement learning for urban industrial emergencies, focusing on production safety accident-induced congestion. We design a Prioritised Experience Replay Deep Q-Network (PERDQN) with an attention mechanism and efficiently model complex urban arterial road networks as a Markov Decision Process. Our approach bases the traffic wave theory alongside BPR functions to evaluate the impact of congestion at different levels of control and assess the overall impact under various traffic control mechanisms. Results of experiments on the Third Ring Road of Beijing show that the proposed method reduces emergency vehicle arrival time by 27.3% compared to traditional algorithms, and 12.8% over standard DQN. The congestion alleviation effectiveness of the method is claimed to be superior in scenarios with more severe traffic, thereby confirming that the method is optimally robust to changing conditions of traffic.