Using Reinforcement Learning for Multi-Objective Cluster-Level NPI Optimization

Xueqiao Peng, Jiaqi Xu, Xi Chen, Dinh Song An Nguyen, Andrew Perrault

Abstract: Non-pharmaceutical interventions (NPIs) play a critical role in the defense against emerging pathogens. Among these interventions, familiar measures such as travel bans, event cancellations, social distancing, curfews, and lockdowns have become integral components of our response strategy. Contact tracing is especially widely adopted. However, the optimization of contact tracing involves navigating various trade-offs, including the simultaneous goals of minimizing virus transmission and reducing costs. Reinforcement learning (RL) techniques provides a promising avenue to model intricate decision-making processes and optimize policies to achieve specific objectives, but even modern deep RL techniques struggle in the high dimensional partially observable problem setting presented by contact tracing. We propose a novel RL approach to optimize a multi-objective infectious disease control policy that combines supervised learning with RL, allowing us to capitalize on the strengths of both techniques. Through extensive experimentation and evaluation, we show that our optimized policy surpasses the performance of five benchmark policies.