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 computational science, researchers utilize Active Subspace Methods (ASM) to identify the directions in a high-dimensional input space that most significantly influence a model's output. This dimension reduction is vital for understanding complex systems. However, a fundamental challenge arises in fields like the social sciences, where Agent-Based Models (ABMs) are prevalent. ABMs simulate individual decision-making, often resulting in discrete or "jumpy" (discontinuous) outputs. Because ASM relies on gradients—the rate of change—it traditionally fails when applied to functions where gradients are either zero or undefined at the point of a jump.
The Problem of Discontinuity and Divergence
A common workaround is to fit smooth "surrogate" models to these jumpy functions. However, this research identifies a critical "divergence pathology": when a smooth interpolant is fitted to a step function, the estimated active subspace can diverge as the sample size increases. Theoretically, ASM asymptotically prioritizes discontinuous directions of variation over continuous ones. In mixed models containing both smooth and jumpy components, the discontinuities eventually "drown out" the smooth trends as more data is collected, potentially leading to erratic sensitivity analyses.
Methodological Extension: Roughness as a Virtue
To address these pathologies, this paper proposes a theoretical bridge. By defining a continuous analog to a regression coefficient ($\beta_r(x)$) that remains stable without differentiability, the active subspace framework is extended to jump-discontinuous functions ($B_{f,\mu}$).
Central to this approach is the selection of the surrogate model. While researchers typically favor "smooth" Gaussian Process (GP) kernels, this study demonstrates that "rougher" kernels, such as Matérn 3/2, are superior for discontinuous functions. These kernels possess higher-order derivatives that are less constrained, allowing them to better capture the active subspaces of non-smooth simulators.
Case Study: The Flee Simulator
The methodology was rigorously tested on Flee, an ABM used to simulate refugee movement across eight real-world displacement crises, including conflicts in South Sudan, Syria, and Ethiopia. By using the estimated active subspace to "prewarp" the input space—stretching important dimensions and shrinking irrelevant ones—the researchers performed K-Nearest Neighbors (KNN) regression with significantly higher accuracy.
Key Findings from Flee:
- Predictive Boost: The truncated active subspace (ASMt) approach improved predictive accuracy by an order of magnitude over baseline methods in several case studies.
- Axis-Alignment: Sensitivity in these social models was surprisingly axis-aligned, meaning individual parameters (like
camp_weightorconflict_weight) dominated the model's behavior rather than complex combinations of inputs. - Domain Insights: The analysis pinpointed specific drivers for different crises; for instance,
camp_weightwas the primary sensitivity driver in Mali and Syria, whileconflict_weightwas most critical in the Central African Republic.
Conclusion and Future Work
This work clarifies the theoretical foundations of applying active subspace methods to discontinuous simulators. By embracing "rougher" surrogates and accounting for the unique behavior of jump-discontinuities, researchers can achieve more stable and accurate sensitivity analyses in the social sciences. Future research may focus on localized ASM to identify exactly where discontinuities occur in black-box simulators, providing a powerful new tool for debugging and refining complex agent-based models.
Summary updated