Development of Evacuation Behavior Simulation Model Using Reinforcement Learning Targeting Immediate and Delayed Storm Surges along the Northern Coast of Kyushu

Abstract

Along the northern coast of Kyushu, storm surges often occur with a significant time lag after a typhoon's passage. This "delayed" characteristic makes it difficult for residents to determine the appropriate evacuation timing. This study aims to develop an evacuation behavior simulation model using Deep Reinforcement Learning to optimize the timing of evacuation decisions.The results revealed distinct behavioral patterns depending on the surge type. For immediate surges, the agent learned "proactive evacuation," triggered by meteorological precursors. In contrast, for delayed surges, the agent tended toward "reactive evacuation," initiating movement only after observing actual water level rises, which significantly reduced the safety margin. Furthermore, the model incorporating astronomical tide levels reproduced "return trips," where agents mistakenly perceived a temporary tidal drop as the end of danger. These findings quantitatively demonstrate the structural risk of escape delays caused by oscillating water levels and suggest the effectiveness of this approach for formulating region-specific disaster prevention strategies.