Neural parameter calibration for large-scale multiagent models.

Area of Science:

• Computational Science
• Quantitative Modeling
• Machine Learning

Background:

• Complex systems in quantitative sciences often require parameter estimation from data.
• Existing parameter estimation methods can be mathematically complex and computationally intensive.
• Models in social sciences, economics, and epidemiology frequently face these challenges.

Purpose of the Study:

• To present a computationally simple and fast method for accurate parameter density estimation.
• To utilize neural differential equations for parameter inference.
• To develop a pipeline combining multiagent models and neural networks for efficient parameter retrieval.

Main Methods:

• Developed a pipeline integrating multiagent models as forward solvers for differential equations.
• Employed a neural network to extract parameters from model-generated data.
• Applied the method to synthetic time series data of the SIR model and the Harris-Wilson economic model.

Main Results:

• The method accurately estimates probability densities for model parameters, even in large systems.
• Demonstrated effectiveness on both synthetic and real-world economic activity data.
• Achieved orders of magnitude greater accuracy and significantly faster computation compared to classical techniques for the Harris-Wilson model.

Conclusions:

• The proposed method offers a computationally efficient and accurate approach to parameter estimation in complex systems.
• Neural differential equations provide a powerful tool for inferring model parameters in diverse scientific domains.
• This approach accelerates the analysis of complex models in fields like epidemiology and economics.