OSSC project, 2023. Authors: Ahmad Hesam (CBS); Frank P Pijpers (CBS); Lukas Breitwieser (ETH Zurich); Peter Hofstee (Delft University of Technology); Zaid Al-Ars (Delft University of Technology).
Calibration is a crucial step in developing agent-based models. Agent-based models are notorious for being difficult to calibrate as they can express various degrees of freedom when model parameters are unknown. Models that appear correctly calibrated to match macro-level observed data perform poorly when micro-level insights need to be inferred. As a result, policymakers cannot be certain that an agent-based model can accurately describe the dynamics of the real-world phenomena that the model tries to mimic. To begin tackling this challenge, we developed a methodology for an epidemiological use case at a full population scale of 17 million agents to observe the effects of using microlevel data for the calibration on the accuracy of the microlevel model outcomes. We show that by calibrating a model on national statistics, but using individual-level microdata, we can on average get 36% more accurate model outcomes on a subnational level. Our model implementation performs two orders of magnitude faster than prior work and allows efficient calibration on HPC computer systems (e.g, ODISSEI Secure Supercomputer).