Surrogate Active Subspaces for Jump-Discontinuous Functions

Nathan Wycoff

Abstract

Surrogate modeling and active subspaces have emerged as powerful paradigms in computational science and engineering. Porting such techniques to computational models in the social sciences brings into sharp relief their limitations in dealing with discontinuous simulators, such as Agent-Based Models (ABMs), which have discrete outputs. Nevertheless, prior applied work has shown that surrogate estimates of active subspaces for such estimators can yield interesting results. But given that active subspaces are defined by way of gradients, it is not clear what quantity is being estimated when this methodology is applied to a discontinuous simulator. We begin this article by showing some pathologies that can arise when conducting such an analysis. This motivates an extension of active subspaces to discontinuous functions, clarifying what is actually being estimated in such analyses. We also conduct numerical experiments on synthetic test functions to compare Gaussian process estimates of active subspaces on continuous and discontinuous functions. Finally, we deploy our methodology on Flee, an agent-based model of refugee movement, yielding novel insights into which parameters of the simulation are most important across 8 displacement crises in Africa and the Middle East.

Introduction

In many scientific domains, particularly the social sciences, researchers rely on complex simulators like Agent-Based Models (ABMs) to understand emergent phenomena such as refugee movement or social unrest. These models often act as "black boxes" with high-dimensional input spaces, making it difficult to identify which parameters drive the simulation's behavior. Global Sensitivity Analysis (GSA), and specifically the Active Subspace Method (ASM), provides a robust framework for identifying these influential directions.

However, ASM traditionally relies on the existence of gradients. Discontinuous simulators, common in ABM research, are either non-differentiable or have zero gradients almost everywhere. While practitioners often fit smooth surrogate models (like Gaussian Processes) to these simulators to estimate gradients, the theoretical justification for what these surrogates are actually measuring has remained elusive. This paper bridges that gap by extending ASM to jump-discontinuous functions.

Theoretical Framework: The Divergence Pathology

A primary challenge in applying smooth surrogates to discontinuous functions is the "divergence pathology." When a smooth surrogate, such as a Gaussian Process (GP), is used to approximate a function with a jump discontinuity, the surrogate must become increasingly steep to bridge the gap as more data points are sampled near the jump. Consequently, the estimated gradient—and thus the active subspace value—diverges to infinity as the sample size increases.

To address this, the paper introduces a theoretical extension of the active subspace for non-smooth functions. This definition is based on the limit of a local linear regression coefficient as the sampling radius approaches zero. This framework clarifies that surrogate ASM is not just a heuristic but a principled way to capture the location and magnitude of discontinuities.

Discontinuity Dominance

A key finding of this research is the principle of "Discontinuity Dominance." In functions containing both smooth variations and sharp jumps, the active subspace estimate will eventually favor the discontinuous directions as the sample size grows. The "jump" creates a surrogate gradient so steep that it mathematically overshadows any smooth trends in the data.

Importance of smooth (X2) vs. discontinuous (X1) directions as a function of sample size.

As illustrated in the synthetic experiments, at small sample sizes, a surrogate might prioritize a smooth trend. However, as more data is collected, the importance of the discontinuous direction (X1) increases while the importance of the smooth direction (X2) diminishes. This trade-off is a fundamental characteristic of surrogate modeling for jumpy systems.

Kernel Selection and Prewarping

The choice of the Gaussian Process kernel is critical for capturing these discontinuities. The study compares several kernels and finds that "rougher" kernels, specifically the Matérn 3/2, outperform smooth Gaussian kernels. The Matérn 3/2 kernel is better equipped to handle the sharp transitions inherent in ABMs without over-smoothing the results.

Once the active subspace is identified, it can be used for "prewarping." This involves transforming the high-dimensional input space into a lower-dimensional subspace that captures the most significant variations. This reduced representation can then be fed into predictive models, such as K-Nearest Neighbors (KNN), to improve efficiency and accuracy.

Case Study: The Flee Simulator

The methodology was deployed on Flee, an ABM designed to simulate forced displacement. The analysis focused on seven parameters across several historical crises, including the South Sudanese civil war, the Syrian civil war, and the Mali coup.

Performance of surrogate active subspaces across six displacement crises, showing error maps and MSE improvements.

The results demonstrated significant predictive gains. By using the active subspace to prewarp inputs, the Mean Squared Error (MSE) of KNN models improved by an order of magnitude in several scenarios, such as South Sudan and Syria. Qualitatively, the analysis revealed that camp_weight (attraction to refugee camps) was the dominant parameter in Mali and Syria, while conflict_weight was more critical in the Central African Republic and Sudan. Furthermore, eigenanalysis showed that most scenarios could be reduced to a 1D or 2D active subspace, significantly simplifying the model's complexity.

Conclusion

This work provides a theoretical foundation for applying gradient-based sensitivity analysis to discontinuous simulators. By understanding how surrogates behave in the presence of jumps, researchers can more effectively use ASM to identify critical parameters and improve the predictive power of complex social science models. Future work will focus on explicitly tuning the balance between smooth and jumpy sensitivity and using surrogate ASM to automatically locate discontinuities in black-box systems.

Summary updated

Abstract

Surrogate modeling and active subspaces have emerged as powerful paradigms in computational science and engineering. Porting such techniques to computational models in the social sciences brings into sharp relief their limitations in dealing with discontinuous simulators, such as Agent-Based Models, which have discrete outputs. Nevertheless, prior applied work has shown that surrogate estimates of active subspaces for such estimators can yield interesting results. But given that active subspaces are defined by way of gradients, it is not clear what quantity is being estimated when this methodology is applied to a discontinuous simulator. We begin this article by showing some pathologies that can arise when conducting such an analysis. This motivates an extension of active subspaces to discontinuous functions, clarifying what is actually being estimated in such analyses. We also conduct numerical experiments on synthetic test functions to compare Gaussian process estimates of active subspaces on continuous and discontinuous functions. Finally, we deploy our methodology on Flee, an agent-based model of refugee movement, yielding novel insights into which parameters of the simulation are most important across 8 displacement crises in Africa and the Middle East.
BibTeX
@inproceedings{wycoff2024surrogate,
  author = {Nathan Wycoff},
  title = {{Surrogate Active Subspaces for Jump-Discontinuous Functions}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2024},
  url = {https://openreview.net/forum?id=uycxGVPVPT},
}
Figure