Reproduce “Neural parameter calibration for large-scale multi-agent models” on OpenResearch

About this paper

Thomas Gaskin, Grigorios A. Pavliotis, Mark Girolami

Researchers from the University of Cambridge, Imperial College London, and The Alan Turing Institute developed a neural parameter calibration pipeline for large-scale multi-agent models, enabling efficient and accurate inference of model parameter probability densities. This method demonstrated two to three orders of magnitude lower prediction error and significantly faster computational speed compared to traditional Bayesian MCMC for diverse model types and real-world datasets.

Key insights

• Probability densities for model parameters can be derived from the loss values computed during neural network training, allowing for comprehensive uncertainty quantification.
• Training the neural network multiple times with different initializations effectively samples the parameter space, enabling identification of multiple minima in non-convex loss landscapes.
• Leveraging modern machine learning libraries' auto-differentiation and vectorized operations streamlines implementation and enhances computational efficiency, provided the numerical solver is compatible.

Summary by alphaXiv.