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

Area of Science:

• Epidemiological Modeling
• Computational Biology
• Data Science

Background:

• Agent-Based Models (ABMs) are increasingly vital for understanding disease dynamics, particularly highlighted during the COVID-19 pandemic.
• Efficient calibration of complex ABMs remains a significant computational challenge, hindering rapid deployment for public health.
• Existing calibration methods often struggle with large parameter spaces and long simulation times.

Purpose of the Study:

• To investigate the efficacy of pruning strategies within a calibration framework for Agent-Based Models (ABMs).
• To evaluate the impact of different pruning techniques on calibration speed and accuracy using a human papillomavirus (HPV) transmission model.
• To provide insights into optimizing ABM calibration for enhanced pandemic preparedness.

Main Methods:

• Developed a novel calibration architecture incorporating pruning techniques.
• Utilized the Optuna framework for integrated calibration of an HPV transmission ABM.
• Simulated six synthetic datasets with varying temporal skewness and tested six pruning algorithms.

Main Results:

• Aggressive pruners excelled with back-heavy datasets, while median pruners were superior for front-heavy datasets.
• Pruning consistently accelerated calibration across all dataset types, often improving or maintaining optimal parameter set accuracy.
• Results were validated using real-world epidemiological data.

Conclusions:

• Pruning is a powerful technique for enhancing the efficiency and effectiveness of ABM calibration.
• This approach offers a cornerstone for improving pandemic preparedness strategies through faster, more accurate epidemiological modeling.
• Further research can explore methods to enhance pruning for balanced datasets.