Reinforcement learning-based tsunami evacuation guidance system

Congestion and crowding are critical issues during indoor and outdoor emergency evacuations. In the 2011 Great East Japan Earthquake and Tsunami, vehicle traffic was one of the causes of the event's casualties. After this, vehicle evacuation in tsunami events is not advised in Japan. Then, pedestrian evacuation is expected to be the primary mode of mobility in emergencies. However, crowding and congestion may affect the evacuation time of individuals and the overall outcome of the process. In addition, narrow streets and a high preference for the shortest routes may worsen the situation. This study aims to find the best evacuation route for a target population, considering less congestion in the road network and increasing the chances of reaching safe areas on time. We propose using reinforcement learning to train an intelligent network of agents, placed at the intersections, in charge of the evacuation process to fully complying evacuee agents. The model rewards decisions that lead to successful evacuation, considering the dynamics of departure times and street congestion throughout the simulation. We demonstrate the applicability of reinforcement learning to guide tsunami evacuation in a simulation and test this against a non-guided case where evacuees move following the shortest paths. Results show that the reinforcement learning model yields better outcomes than evacuations following the shortest paths.