John Nardini, Department of Mathematics and Statistics, The College of New Jersey

Decoding agent-based model behavior: novel methods for prediction and global sensitivity analysis

Agent-based models (ABMs) are invaluable tools for studying the emergence of collective behavior in biology. Unfortunately, it is challenging to analyze ABM behavior due to their computational and stochastic nature. In this talk, I will present two recent studies aimed at developing new methodologies to enable the prediction, interpretation, and analysis of ABMs. In the first study, we use biologically-informed neural networks (BINNs) to forecast and predict ABM behavior. In particular, we show BINNs can learn interpretable differential equations to predict ABM data at new parameter values, and demonstrate this success using three case study ABMs of collective migration. In the second study, we combine several machine learning algorithms to develop a global sensitivity analysis pipeline for ABMs that is capable of identifying sensitive parameters, revealing common model patterns, and linking input model parameters to these patterns using a spatial ABM of tumor spheroid growth. Taken together, these studies demonstrate how concepts from machine learning are valuable for studying ABMs and will advance data-driven ABM modeling.