Optimizing evacuation paths using agent-based evacuation simulations and reinforcement learning

Evacuation path optimization during major flooding events is crucial for minimizing casualties. Notably, recent studies have underscored the importance of considering multiple factors, such as inundation timing, road congestion, and evacuation destination capacities, during path optimization for effective flood evacuation planning. Drawing insights from these studies, the present study developed a novel methodology to optimize evacuation paths for individual evacuees by integrating agent-based tsunami evacuation simulations with Q-learning, a well-known reinforcement learning technique . The effectiveness of the proposed methodology was tested in a tsunami-prone coastal area . Furthermore, to comprehensively assess the performance of the methodology under varying conditions, several scenarios with diverse reward settings and evacuation start times (5, 10, and 15 min after the earthquake) were simulated. The results demonstrated that the proposed methodology significantly reduced the number of casualties by dispersing evacuees across wide areas, alleviating road congestion, and guiding evacuees toward evacuation destinations with adequate capacity. Notably, when rewards for reaching evacuation destinations were set significantly higher than typical inundation times, and differences in inundation onset times between nodes were integrated into reward calculations, the proposed methodology achieved mortality rate reductions of approximately 60 % compared to the traditional shortest-path methodology.