Computing Optimized Epidemic Control Policies Using Reinforcement Learning

Abstract

This paper presents a reinforcement learning-based algorithm for computing epidemic contention policies defined interms of mobility-restriction actions. The algorithm’s objective is simultaneously minimizing public health and economic affectations, which is challenging because both objectives are in conflict. We used a SEIRD (Susceptible-Exposed-Infected-Recovered-Deceased) epidemiological model to capture the spreading dynamics of a disease characterized by the probabilities of transitioning between the states defined in the model. To train the reinforcement learning algorithm, we implemented a discrete event simulator from scratch that considers different mobility patterns and diseases defined in terms of the SEIRD model probabilities. Extensive simulation-based results show that the proposed algorithm computes mobility restriction policies that effectively minimize the two opposite objectives and are flexible enough to allow a decision-maker to prioritize either public health or the economy.