Scaling Agent-based Epidemic Diffusion on HPC Clusters

Abstract

Over the course of the COVID-19 pandemic, a wide variety of modeling approaches have been employed to inform the decisions made by policymakers at various levels of government. This process has helped to illuminate both the benefits and the limitations of using different modeling techniques in this role. Agent-based models (ABMs) – where the behaviors of individual members of a population are simulated directly – have proved particularly well-suited in use cases like counterfactual analysis. Counterfactual analysis seeks to understand the impact of different sets of public health interventions in various what-if scenarios, which are often easier to directly represent in ABMs. However, this resolution comes at a cost; ABMs are generally orders of magnitude more complex and computationally expensive than other modeling techniques. Thisgenerally means that such ABMs must be highly scalable parallel applications. Existing ABMs are generally either (1) complex, small simulations (at most around a million agents) with a focus on epidemiological results rather than computational efficiency, or (2) simpler, large models where much of the complexity lies in the underlying datasets and the behavioral model of agents remains relatively simple (e.g. a sequence of top-down interventions determines behavior). With these limitations of existing simulations in mind, we propose to enable ABMs of infectious disease spread to efficiently scale to large populations and core counts while efficiently modeling a combination of top-down and bottom-up behaviors that are both complex and dynamic. This work involves two main directions: (1) increasing the scalability of ABMs for large populations of agents and (2) introducing more complex behavioral models into large-scale ABMs, particularly ones which dynamically spread and co-evolve with the simulated disease.