Surrogate assisted strategies for the parameterisation of infectious disease agent–based models

Abstract

Agent-based modelling and simulation (ABMS) is a viable solution for real-time decision analysis and policy-making towards preventing an epidemic. This is due to its ability to model the real world by incorporating more complexity than previously used compartmental models. Model validation, parameter calibration and long simulation time are significant limitations of ABMS. Surrogate models (SMs) can overcome these limitations. However, there is a lack of comprehensive comparison between surrogate assisted parameterisation strategies. In addition, there is a lack of research on using an SM to tackle problems outside of parameterisation. We provide a comparison of some state-of-the-art and classical intelligent sampling, optimisation and evolutionary methods for parameter calibration along with a framework for evaluation of these methods. The extensive experimental results show that the Dynamic Coordinate Search Using Response Surface Models paired with the XGBoost SM outperforms competing methods regarding accuracy and speedup achieved on synthetic epidemic data. Parameterising an ABM taking an average similarity score across each output distribution allows the parameterisation approach to more closely match real-world data. Lastly, we have shown that a Long Short-Term Memory network SM can replicate the transmission dynamics of acomplex ABM, significantly reducing simulation time whilst maintaining accuracy.