Road-reconstruction after multi-locational flooding in multi-agent deep RL with the consideration of human mobility - Case study: Western Japan flooding in 2018 -

Abstract

Record-breaking heavy rain occurred in Western Japan from June 28 to July 8, 2018. Many roads in Hiroshima and Okayama Prefecture were disrupted simultaneously. The government desired the rapid recovery of disrupted human mobility; however, the restoration of roads that were frequently used by citizens was delayed for a week after this flooding. There are three challenges to overcome for an efficient disaster road management plan: 1) the lack of prior knowledge, 2) the absence of evaluation indicators, and 3) the non-consideration of human mobility changes post-disaster. To address these limitations, we developed data-driven reinforcement learning modeling of human movement in the recovery operation by utilizing mobile phone GPS data. The input data and reward were defined by the damage recovery process, the changes in road usage with the restoration, and the travel time. In this model, the agents learned the restoration effect of each target through interaction and feedback, and subsequently restored the roads preferentially with a high effect of human mobility recovery. Furthermore, they could learn the cooperation with numerical protocol regarding road usage in 1000 types of origin-destination pairs. The visualization of the road system may enable administrations to provide additional responses to alleviate abnormal mobility. Accordingly, with a generalized analytical framework, the government would acquire a more efficient road management plan post-disaster with the consideration of human mobility.