Forecasting and Predicting Stochastic Agent-Based Model Data with Biologically-Informed Neural Networks

by John T Nardini

Read the paper

Agent-based models (ABMs) are widely used to study biological systems, but heavy computational requirements limit our ability to predict their behavior. Differential equation (DE) models are often used as ABM surrogates, but they can provide poor predictions. We propose that biologically-informed neural networks (BINNs) can learn informative DE models that predict ABM behavior. We demonstrate how BINNs’ learned DE models can forecast future ABM data at new parameter values. We highlight the strong performance of this methodology in three case study ABMs that explore different rules on cell-cell interactions in collective migration. BINNs learn predictive and interpretable DE models even when other DE models are ill-posed or complex.

1) We explore several different ABM rules and summarize the ABM density over time. 2) BINN models can be trained to the ABM data. 3) We predict new ABM data by simulating the BINN's learned PDE model.