Route optimization using deep reinforcement learning with simulation-based data generation: a case study of evacuation guidance in a radiological emergency

Abstract

The route planning problems have been successfully addressed by reinforcement learning (RL) techniques. However, it is a quite costly task for disaster researchers or policymakers to solve their specific cases with RL. Recently, machine training with simulation-based data generation has been suggested to cut down the cost of RL applications with the benefits of straightforward visualization and ease-of-use. This research demonstrates a route optimization for evacuation guidance by using the machine training strategy. A multi-agent system is built to simulate evacuation dynamics of people on a road network, near to Kori nuclear power plant in South Korea. There are two types of agents. One is in charge of an evacuee who has preset behavioral rules for individual responses, whereas the other tests and advances its routing strategy for evacuation guidance. In the different scenarios, related to social interactions on street and social networks, the training agent try to minimize the evacuation time of residents. The results indicate that the simulation-based strategy can provide a well-trained artificial intelligence to the specific case of evacuation guidance in a radiological emergency to improve the averaged evacuation dynamics. Therefore, this study contributes to practical applications of RL which have been barely incorporated in radiological emergencies.