Modelling the behaviour of corporations during the flood damage recovery process using multi‐agent deep reinforcement learning

Abstract

Abstract In the event of a flood, affected firms will experience production stoppages, resulting in supply chain (SC) disruptions and indirect economic damage to firms even outside the affected area. To mitigate the economic damage, it is essential for each firm in the entire SC to acquire a behavioural strategy in line with its recovery status. To achieve optimal behavioural strategies for individual firms in the recovery process, we present a learning framework using multi‐agent simulation and deep reinforcement learning, which considers dynamically changing factors such as the flooding depth of buildings and roads, SC, labour and capital. The agents use reinforcement learning to make decisions about the dynamically changing environment and establish optimal action strategies during the recovery process. The model is trained using inundation simulation data from the Arakawa River located in Tokyo and inter‐firm transaction data in Japan. The results show that the learned model can acquire behavioural strategies in the recovery process on a firm‐by‐firm basis, and to identify industries and regions with high economic damage. The results of this study can contribute to the development of recovery plans that consider the early recovery of SCs during floods.

Keywords

• deep reinforcement learning
• flood
• inter‐firm transaction data
• multi‐agent simulation
• supply chain disruption