Reinforcement learning approach for adapting complex agent-based model of evacuation to fast linear model

Nowadays, most coastal regions face a potential risk of tsunami. The evacuation is one of the most effective mitigation procedures. However, there are always the part of evacuees (e.g. the tourist) who lack information of the evacuation map, which motivates us to focus on the problem of optimizing of guidance sign placement for tsunami evacuation. Concretely, we must find out an optimal sign placement in order to have as many evacuees as possible to reach shelters before tsunami arrival. In fact, most studies focus on two approaches: Agent-Based modeling and Equation-Based modeling. Each approach has its own pros and cons. While the Agent-Based modeling introduces an accurate but very slow model, the Equation-Based one provides a very fast but inaccurate model. The idea of this study is that we learn the accurate Agent-Based model and adapt it into very fast Equation-Based model in order to solve the optimizing problem. In this paper, we present clearly two models representing the two approaches and pros and cons of each model. We then propose a reinforcement learning approach for adapting complex Agent-Based model into a very fast Linear Model (representing Equation-Based modeling approach). By experimentation, our proposed approach shows that we can replace a slow complex model by a very fast model with an acceptable level of accuracy in order to solve optimizing problem.