Speaker
Description
Agent-based models (ABMs) capture heterogeneous contacts, stochastic transmission, and complex interventions in infectious disease dynamics but are computationally expensive, limiting parameter inference and policy analysis. We develop an equation-learning framework that derives interpretable ordinary differential equation (ODE) surrogates directly from stochastic ABM simulations. Using the COVID-19 ABM Covasim, we construct a 12-compartment ODE system incorporating testing and contact tracing while treating key epidemiological rates as state-dependent functions rather than constants. These functions are learned from ABM-generated trajectories using Biologically Informed Neural Networks (BINNs), which denoise state dynamics while enforcing epidemiological constraints. Sparse regression is then applied to obtain symbolic representations of the learned parameter functions to enable model interpretability. To quantify uncertainty arising from ABM stochasticity, we integrate Approximate Bayesian Computation to infer posterior distributions over model coefficients. The resulting ODE surrogates accurately reproduce ABM dynamics under constant and time-varying interventions and provide a fast, interpretable framework for analyzing complex epidemic models.
