Using Reinforcement Learning Algorithms to Explore COVID-19 Spread in South Africa

Hanie, R. L.; Van Rensburg, J. T. J.

Many learning opportunities for machine learning (ML) exist within the context of how viruses and their spread can be combatted. If the agents can be trained to demonstrate optimal behaviour in a pandemic, their actions can possibly be replicated to improve spread in a real-life scenario. The aim of this research is to train reinforcement learning (RL) agents to survive in a rule-based AI environment that simulates the spread of COVID-19 in South Africa. The RL agents used in the training environment were created using Unity's ML agent SDK. The ML agent SDK supports the usage of two RL-specific algorithms, Proximal Policy Optimization (PPO), and Soft Actor-Critic (SAC). This study contributes to the AI space by providing insight into how a virus and its interaction with a population can be modelled using Unity and machine learning. The agents were able to combat COVID-19 effectively and did so by self-Training how to maintain social distance and have regular check-ups at the hospital. It was also observed that susceptible agents pay frequent visits to the hospital without ever being rewarded for doing so. The code will be open-sourced to the Unity machine learning agent's SDK community and discords. © 2021 IEEE.