Firm-level behavior control after large-scale urban flooding using multi-agent deep reinforcement learning

With natural disasters have become large scale, diversified, and frequent, the indirect economic damage due to interruption of supply chain tends to be large. Therefore, it is important to recover as quickly as possible for companies after disasters. In this paper, we use reinforcement learning to optimize a company's action strategy so that it can efficiently recover the inter-firm transaction network in the supply chain after large-scale urban flooding. The agent holds information on disaster and supply chains obtained from inter-firm transaction data and flood simulation analysis data, enabling us to create a simulation with detailed urban infrastructure information by using the high-dimensional data to construct detailed spatial states. The paper also proposes an action policy for companies based on multi-agent deep reinforcement learning to optimize the behavior of companies in the recovery process. This work bridges the divide between high-dimensional data set input and post-disaster behaviors, enabling an artificial agent to learn the best action to take after a disaster. Our results are as follows. Through learning, agents can recover efficiently after a disaster. Companies tend to secure alternative business partners first and then perform recovery work and business expansion.