Can pruning improve agent-based models' calibration? An application to HPVsim

Agent-Based Models (ABMs) have gained popularity over the COVID-19 epidemic, but their efficient calibration remains challenging. Here we propose a novel calibration architecture by investigating the role of pruning in ABM calibration. We use a recently developed model for human papillomavirus (HPV) transmission and focus on its integrated calibration framework, Optuna. Simulating six synthetic datasets of various temporal skewness, with six pruners, we show that more aggressive pruners perform best (in terms of loss function at end of calibration) for very-back-heavy datasets, while median pruners are better for more-front-heavy datasets. For more balanced datasets most of the pruners perform similarly to no pruning. However, across all datasets pruning notably sped up calibration, in many cases without compromising on - or even improving upon - the optimal found parameter set. We validate our results through application to real-life data. Finally, we discuss approaches for improving "bad pruners" for balanced datasets. Our proof-of-principle study shows that pruners can improve ABMs' calibration. As ABMs are becoming more widely used in epidemiological modelling, designing the next level of pandemic preparedness strategies will need to address efficient calibration; we believe pruning is a cornerstone for this.