Machine learning assisted calibration of stochastic agent-based models for pandemic outbreak analysis

Abstract

Mathematical modelling with agent-based models (ABMs) has gained popularity during the COVID-19 pandemic, but their complexity makes efficient and robust calibration to data challenging. We propose an improved method for calibrating ABMs that combines a machine-learning step with Approximate Bayesian Computation (ML-ABC). We showcase its application to Covasim - a stochastic ABM that has been timely and responsively used to model the English COVID-19 epidemic and inform policy at important junctions. We illustrate the advantage of ML-ABC application in calibrating Covasim during the first and the second COVID-19 epidemic waves of 2020 and early 2021, demonstrating that the use of an ML screening step allows us to derive faster and more efficient estimates of the posterior distribution of the Covasim optimal parameters without compromising on accuracy. This is important for generating timely responsive modelling results during an emerging epidemic.