Neural parameter calibration for large-scale multi-agent models

Thomas Gaskin, Grigorios A. Pavliotis, Mark Girolami

Introduction

Computational models are indispensable for policy-making, from tracking epidemiological outbreaks like COVID-19 to simulating complex urban economic activities. However, these models frequently depend on "free parameters"—such as infection rates or consumer sensitivities—that are not directly observable. Calibrating these parameters typically involves comparing model outputs to real-world data using statistical methods like Markov Chain Monte Carlo (MCMC).

The primary challenge is computational: MCMC often requires thousands of simulations to converge, creating a massive bottleneck for large-scale systems with numerous agents. Furthermore, many calibration landscapes are non-convex, leading traditional methods to stall in local minima. This paper introduces a high-performance Neural Parameter Calibration (NPC) pipeline that leverages neural networks to predict model parameters directly from data, achieving a 200–400x speedup over traditional Bayesian methods while significantly improving accuracy.

Methodology: Physics-in-the-Loop

The core innovation of this approach is treating a neural network not as the model itself, but as a sophisticated "parameter extractor."

The Neural Parameter Calibration pipeline integrates a neural network with a numerical solver, allowing for gradient backpropagation through physical equations.

As illustrated in the pipeline above, the process follows these steps:

  1. Parameter Extraction: A neural network ($u_\theta$) receives observed data $\mathbf{T}$ as input and outputs predicted parameters $\hat{\lambda}$.
  2. Numerical Simulation: These parameters are passed to a Numerical Solver (e.g., an Agent-Based Model or ODE solver), which generates predicted data $\hat{\mathbf{T}}$.
  3. Differentiable Training: A loss functional $J$ compares $\hat{\mathbf{T}}$ to $\mathbf{T}$. Because the solver is integrated into the training loop, gradients $\nabla_\theta J$ can be backpropagated through the physical equations, effectively a "Neural ODE" approach for parameter estimation.
  4. Density Estimation: Beyond point estimates, the authors utilize the training trajectory to estimate the full probability density (posterior) of the parameters. By treating the loss at each step as a potential, they achieve Bayesian-like uncertainty quantification without the high cost of sampling.

Experimental Validation and Results

The NPC pipeline was validated across three diverse scenarios: a synthetic SIR epidemiological model, a Harris-Wilson economic model, and real-world economic data from Greater London.

Performance Gains

The results demonstrate a transformative shift in calibration efficiency. For the Greater London dataset—which incorporates travel convenience metrics from the Google Distance Matrix API—the NPC method produced results in approximately one minute. In contrast, state-of-the-art MCMC approaches required between 3 and 7 hours. This represents a speedup of roughly 195 to 390 times. Furthermore, the prediction error was reduced by two to three orders of magnitude compared to classical techniques.

Calibration of the SIR model. The top panel shows ground truth trajectories, while the bottom panel shows the NPC predictions with shaded uncertainty regions.

Scalability and Identifiability

The method scales quadratically with the number of system components (e.g., origin and destination zones), making it viable for large-scale urban simulations. A key insight from the study is the relationship between data noise and parameter identifiability. As noise increases, the "non-trivial" parameter solutions become less distinct, providing a visual diagnostic for when a model's parameters can no longer be reliably recovered from available data.

Discussion: Hybrid Intelligence

This research advocates for a "hybrid intelligence" approach. It preserves the "white-box" integrity of differential equations and physical laws that scientists trust, while employing "black-box" neural networks to navigate high-dimensional, non-convex parameter spaces efficiently. This synergy allows the training process itself to act as a sampler, providing insights into the likelihood surface and identifying multiple global minima—a common feature in complex economic systems.

Limitations and Future Outlook

While highly effective, the authors note several areas for future development:

  • High-Dimensional Parameters: The current testing focused on global parameters; scaling to learn individual edge weights in massive social networks remains a challenge.
  • Solver Engineering: The pipeline requires fast, auto-differentiable numerical solvers (e.g., in PyTorch), which demand significant engineering effort to implement.
  • Numerical Stability: Neural networks may occasionally predict parameters that lead to chaotic or unstable regimes in the physical model. The authors suggest using constrained activation functions (like sigmoids) to keep predictions within physically realistic bounds.

Summary updated

BibTeX
@inproceedings{gaskin2023neural,
  author = {Thomas Gaskin and Grigorios A. Pavliotis and Mark Girolami},
  title = {{Neural parameter calibration for large-scale multi-agent models}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2023},
  doi = {10.1073/pnas.2216415120},
}
Figure