[4Xin1] Poster session 2

[4Xin1-56] Prediction of Resident Evacuation during Flooding Using Multi-Agent Simulation with Reinforcement Learning

〇Tsukasa Yoshida¹, Kazuki Okumura¹, Yutaka Arakawa¹, Narichika Nomoto¹, Yasuhiro Iida¹, Takao Nakamura¹ (1.Nippon Telegraph and Telephone (NTT))

Keywords:Multiagent Simulation, Reinforcement Learning, Disaster Evacuation

It is important to improve simulation performance since floods that cause enormous damage have been increasing in recent years. In previous studies of multi-agent simulation, the destination of the agents was generally fixed and their actions were restricted. As a result, it was not possible to reproduce the movements that humans could take in the event of a real disaster, resulting in a discrepancy with actual human behavior. To solve this problem, we propose a new paradigm for evacuation forecasting which regarded human behavior as a resource recovery game, and to learn it by reinforcement learning. Additionally, we propose reinforcement learning using Monte Carlo Tree Search (MCTS) as a method for solving this game. In the evaluation experiments, the MCTS algorithm was able to learn the correct behavior, and reproduced the situation where some people stayed in the flooded area like the actual behavior.