A computationally efficient workflow for calibrating expensive, high-resolution policy-oriented agent-based models. Several rounds of history matching use a heteroskedastic Gaussian-process emulator as a cheap surrogate for the simulator to iteratively rule out implausible regions of the parameter space; the surviving plausible region is then calibrated with approximate Bayesian computation (ABC). Demonstrated on Covasim, a widely used COVID-19 agent-based model, by matching time series of diagnoses and deaths across random seeds, the method sharply reduces the compute of the original hyperparameter-optimization calibration (over 100,000 runs, ~35 days) while improving uncertainty quantification, widening the range of policy models for which timely calibration is feasible.
Presents a more computationally efficient way to calibrate high-resolution, policy-oriented agent-based models, whose expensive simulation runs make timely calibration and policy analysis difficult. …