Using machine learning as a surrogate model for agent-based simulations.

Area of Science:

• Computational Science
• Artificial Intelligence
• Complex Systems Modeling

Background:

• Agent-based models (ABMs) are powerful tools for simulating complex systems.
• Analyzing ABM outputs is computationally intensive due to non-linear parameter relationships and long run times.
• Traditional Monte Carlo methods are often too costly for comprehensive ABM analysis.

Purpose of the Study:

• To evaluate machine learning (ML) methods as surrogate models for ABM analysis.
• To compare the performance of various ML techniques against established surrogate modeling approaches.
• To identify optimal ML strategies for efficient and accurate ABM surrogate modeling.

Main Methods:

• Proof-of-concept evaluation of multiple ML algorithms (e.g., ANNs, gradient-boosted trees, Gaussian processes).
• Comparative analysis of surrogate model performance in replicating complex ABM behaviors.
• Assessment of accuracy and computational time across different ML methods.

Main Results:

• Artificial neural networks (ANNs) and gradient-boosted trees generally outperform Gaussian process surrogates for ABM surrogate modeling.
• ANNs demonstrated the highest accuracy in replicating model behavior with numerous runs, despite longer training times.
• ML-based surrogate models offer a viable alternative to computationally expensive methods for ABM analysis.

Conclusions:

• Machine learning methods, particularly ANNs, provide a robust and efficient approach for surrogate modeling of agent-based models.
• Implementing ML for surrogate modeling can significantly reduce CPU time for ABM calibration and analysis.
• This approach facilitates more thorough sensitivity analyses, leading to deeper insights into complex system dynamics.