Using Control Theory and Bayesian Reinforcement Learning for Policy Management in Pandemic Situations

As engineers and scientists, it is our responsibility to learn lessons from the recent pandemic outbreak and see how public health policies can be effectively managed to reduce the severe loss of lives and minimize the impact on people’s livelihood. Non-pharmaceutical interventions, such as in-place sheltering and social distancing, are typically introduced to slow the spread (flatten the curve) and reverse the growth of the virus. However, such approaches have the unintended consequences of causing economic activities to plummet and bringing local businesses to a standstill, thereby putting millions of jobs at risk. City administrators have generally resorted to an open loop, belief-based decision-making process, thereby struggling to manage (identify and enforce) timely and optimal policies. To overcome this challenge, this position paper explores a systematically designed, feedback-based strategy, to modulate parameters that control suppression and mitigation. Our work leverages advances in Bayesian Reinforcement Learning algorithms and known techniques in control theory, to stabilize and diminish the rate of propagation in pandemic situations. This paper discusses how offline exploitation using pre-trigger data, online exploration using observations from the environment, and a careful orchestration between the two using granular control of multiple on-off control signals can be used to modulate policy enforcement based on established metrics, such as reproduction number.