Simulating multi?exit evacuation using deep reinforcement learning

Conventional simulations on multi-exit indoor evacuation focus primarily how to determine a reasonable exit based numerous factors in changing environment. Results commonly include some congested and other under-utilized exits, especially with large numbers of pedestrians. We propose simulation deep reinforcement learning (DRL), referred as the MultiExit-DRL, which involves neural network (DNN) framework facilitate state-to-action mapping. The DNN applies Rainbow Deep Q-Network (DQN), DRL algorithm that integrates several advanced DQN methods, improve data utilization stability further divides action space into eight isometric directions for possible pedestrian choices. compare MultiExit-DRL two conventional models three separate scenarios: varying distribution ratios; width open schedules an exit. results show presents great efficiency while reducing total number frames all designed experiments. In addition, integration allows pedestrians explore potential exits helps optimal directions, leading high utilization.