Simulating COVID-19 contagion patterns using a machine-learning-augmented agent-based model

Agent behavior in simulation models is often mechanistic and lacking in realism. This chapter shows how machine learning (ML) techniques can be harnessed to increase behavioral realism in agent-based models (ABM). We use an inverse reinforcement learning (IRL) algorithm to derive the decision rules governing agent actions in the context of mobility in cities under COVID-19 restrictions. We combine the trained IRL algorithm with a pre-existing spatial epidemiological ABM that simulates agent behavior for real-world environments at the building level. These feed into the ABM, generating agent mobility trajectories using a Monte Carlo Markov Chain (MCMC) process. This is illustrated using a case study of COVID-19 contagion in Jerusalem city center. Given the level of spatial granularity in this approach, simulations of COVID-19 mitigation measures are feasible for different urban scales (city block, neighborhood, central business district, and so on). The chapter outlines the future challenges for generating behaviorally enhanced agent-based simulations.