REINFORCEMENT LEARNING ANALYSIS OF THE IMPACT OF FLOOD EXPERIENCE ON DECISION MAKING CRITERIA OF EVACUATION BEHAVIOR

Abstract

 In order to analyze the effect of flood experience on the decision-making criteria for residents' evacuation behavior, an agent model that simulates the acquisition of evacuation criteria by reinforcement learning from the flood experience simulated on a computer is proposed. The model repeats the experience where it starts to evacuate with reference to nearby river water level as an evacuation inducing switch and receives reward according to the appropriateness of its behavior. It obtains the evacuation criteria (threshold value of river water stage) which maximize the reward value. Through the learning process based on various rewarding system and various scales of virtual flood experiences, the most promising rewarding rule is to give positive value only when the agent reaches the shelter safely in flood cases where its house is flooded.