Accelerating Multiagent Epidemic Modeling with Surrogate-Based Methods

In this study, we explore the potential of machine learning models as surrogates for computationally intensive probabilistic models. Traditional multi-agent models (MAM) for forecasting epidemic outbreaks, specifically acute respiratory infections, are accurate but require significant computational resources and time. We propose a surrogate Auto Encoder (AE) to approximate the outputs of these MAMs efficiently. The results indicate that the AE can effectively approximate the MAM outputs, significantly reducing the computation time while maintaining acceptable accuracy. This surrogate model serves as a preliminary assessment tool to decide whether running the full MAM is necessary, thereby saving resources. However, the AE struggles with inheriting the stochasticity of the MAM, resulting in minimal variation in predictions for identical inputs. Despite these limitations, our findings suggest that machine learning surrogates hold promise for accelerating multiagent epidemic modeling, which enhances the applicability of the latter for the epidemic surveillance.