Smart Policy Design for COVID-19: A Deep Reinforcement Learning Framework

Abstract

More than 700 million people became infected and about 7 million have died from COVID-19, making it one of the deadliest pandemics in history. It triggered severe economic recession around the world that disrupted multiple industries such as agriculture, manufacturing and tourism. Governments around the world responded by implementing interventions to control disease transmission, but their impact varied from one country to another. There have been numerous studies to understand the effectiveness of the interventions but there are considerable variations in the interpretation. However, it is evident that tailoring policies for individual regions makes them more impactful. In the research, we decided to focus on regions of the United States and treat each state individually because of their diversity. We employ deep reinforcement learning to handle the dynamic nature of the disease, specifically multi-agent reinforcement learning where individual agent handles distinct states in a shared space to mimic realistic environment. The agent intervention differed from the actual intervention most of the time. Overall suggestion is to implement intervention earlier with higher intensity then gradually reduce aggressiveness and maintain a moderate level of intensity. However, this will depend on how much we want to prioritize public health over national economy.