What is it about?
We calibrate the parameters of mathematical models of complex systems to real data using a neural network. The neural network achieves this task in seconds where more classical statistical methods take hours, while the quality of the calibration is significantly improved. Moreover we can exploit the neural network's training process to give uncertainty quantification on the parameters.
Featured Image
Photo by Erick Butler on Unsplash
Why is it important?
This method will considerably simplify and improve mathematical models of complex systems in the computational sciences, including economics, social sciences, computational epidemiology, and of course applied mathematics. Our method is simple to implement, robust, works for a variety of different datasets and problem types (including convex and non-convex, at least in low dimensions), and can make use of powerful machine learning libraries.
Perspectives
We hope this article will help researchers across the quantitive sciences improve their model calibrations. We are currently expanding the method to further problems, e.g. high-dimensional parameters such as networks, and are excited to explore the many lines of research arising from this work.
Thomas Gaskin
University of Cambridge
Read the Original
This page is a summary of: Neural parameter calibration for large-scale multiagent models, Proceedings of the National Academy of Sciences, February 2023, Proceedings of the National Academy of Sciences,
DOI: 10.1073/pnas.2216415120.
You can read the full text:
Resources
Contributors
The following have contributed to this page
