From Model-Based Screening to Data-Driven Surrogates: A Multi-Stage Workflow for Exploring Stochastic Agent-Based Models

A team of researchers has published a paper outlining a novel, automated pipeline designed to tackle the significant challenges of analyzing complex Agent-Based Models (ABMs). These models, which simulate the actions and interactions of autonomous agents, are powerful tools in fields like economics, epidemiology, and ecology. However, their inherent stochasticity and high dimensionality—often called the 'curse of dimensionality'—make systematic exploration and sensitivity analysis extremely difficult and computationally expensive. The new workflow directly addresses this by integrating systematic design of experiments with machine learning.

The methodology proceeds in two key stages. First, an automated model-based screening process runs the ABM to identify the most influential variables, assess outcome variability, and intelligently segment the vast parameter space into more manageable regions. Second, the team trains Machine Learning models to act as fast, data-driven surrogates, learning to map the complex, nonlinear interaction effects between the remaining variables. This hybrid approach automates the discovery of 'unstable regions' where system outcomes are highly sensitive to specific parameter combinations. Demonstrated on a classic predator-prey case study, the framework provides modelers with a rigorous, largely hands-off tool for performing sensitivity analysis and testing potential policies, even when dealing with simulators that have thousands of uncertain parameters.