Source-linked AI summary
Efficient regime boundary detection in discontinuous simulation models
Federico Betti, Eleni Chatzi, Paolo Tiso
TL;DR
Discontinuous or non-smooth simulator outputs make regression-based level-set estimation a difficult route to regime boundaries. The paper instead uses SVM classification and continuous tangential boundary exploration; it reports improved sample efficiency over active-learning alternatives across analytical and simulation examples.
Problem
For discontinuous or non-smooth simulator outputs, regression-based level-set estimation can misrepresent the boundary, and accurate boundary detection is needed to distinguish regimes.
Source: §1 · p. 1, §3.3 · p. 5Method
The approach classifies regimes with an SVM and uses continuous Hamiltonian Monte Carlo exploration of the estimated boundary's tangential directions, followed by space-filling downselection.
Source: §1 · p. 2Results
The authors report better sample efficiency than uncertainty-based sampling and overly diffusive gradient exploration on analytical examples, with stronger gains for highly curved and higher-dimensional boundaries.
Source: §1 · p. 2, §6 · p. 25Takeaways & Limitations
Boundary detection can support partitioning parameter space for regime-specific localized surrogate or reduced-order models.
Source: §1 · p. 1HMC's numerical bias vanishes under stated conditions summarized by strong convexity, which the authors note does not hold for Gaussian-kernel combinations and requires regularization even for a linear SVM.
Source: §4.2 · p. 10
Abstract
from arXiv · showhide
Many computational models exhibit multiple operating regimes separated by abrupt or non-smooth transitions as the input parameters vary. Such regime transitions complicate the construction of global parametric reduced-order models, often necessitating large training datasets to capture the distinct system behaviors. Accurate regime boundary detection instead enables localized modeling strategies by partitioning the parameter space into regions of qualitatively distinct behavior. In many cases, regime switching corresponds to the excursion set of a simulator output. When such a quantity is discontinuous or non-smooth, regression-based algorithms for level set estimation can struggle to correctly approximate the boundary. In this work, we show that solving a resulting binary classification problem can improve sample efficiency and we further propose a novel adaptive sampling strategy for support vector machines. Unlike uncertainty-based sampling schemes, the proposed approach uses a continuous search space and efficiently explores the currently estimated boundary in its tangential directions by gradient-based dynamic sampling. Because the decision function admits analytical gradients, this exploration incurs negligible computational overhead relative to the underlying model evaluations. Then, a space-filling subset of the generated samples is selected as the next batch to be queried. We demonstrate improved sample efficiency compared to existing active learning algorithms on analytical and numerical examples. The proposed method provides a sample-efficient framework for regime boundary detection in a broad class of discontinuous simulation-driven problems.
1. Introduction
Abrupt regime changes complicate global reduced-order modeling, while accurate boundary estimates can support localized models and reveal system behavior. This work frames boundary detection as classification and proposes continuous, tangential boundary exploration to improve sample efficiency.
Motivation
- Global reduced-order models can be inefficient across abrupt regime changes, while identifying the boundary supports localized models and can inform safety assessment or reliability analysis.Large principal angles can ill-condition manifold interpolation, and a global POD basis may require many modes.Source: §1 · p. 1
Boundary detection challenge
- Because the separating surface is usually unknown and exhaustive sampling is infeasible for expensive evaluations, adaptive sampling must refine it from limited queries.New evaluations should target informative regions rather than rely only on static space-filling designs.Source: §1 · pp. 1–2
- For discontinuous or non-smooth outputs, regression may mislocate the boundary or fail because the excursion-set boundary need not coincide with the output's threshold level set.Smooth-kernel Gaussian processes can over-smooth discontinuities and sharp variations.Source: §1 · p. 2
Contribution
- The method casts regime identification as binary classification with an SVM, then explores its estimated boundary tangentially in a continuous space using Hamiltonian Monte Carlo and selects a space-filling query subset.The authors report increasingly pronounced sample-efficiency gains for more curved boundaries and higher-dimensional parameter spaces.Source: §1 · p. 2
2. Background
The problem is to approximate a regime boundary with few evaluations of a deterministic, potentially discontinuous simulator. The section motivates classification and adaptive boundary sampling as a route to regime-aware localized models.
2.1. Problem statement
- The problem assumes a bounded parameter domain, a deterministic black-box output that may be non-smooth or discontinuous, and a threshold marking a fundamental change in system behavior.The stated goal is an accurate boundary approximation with as few simulator queries as possible.Source: §2.1 · pp. 2–3
2.2. Adaptive sampling and localized modeling
- Prior adaptive and localized modeling approaches include partitioning, surrogate enrichment, and reduced-order-model sampling, but some scale poorly, lack boundary targeting, or use clusters that need not match physical regimes.The review also notes that static discontinuity detection with localized Gaussian processes does not adaptively sample to locate the discontinuity.Source: §2.2 · p. 3
- The proposed framework identifies regimes for later construction of tailored local models and directly classifies parameter points instead of regressing a discontinuous output to recover its threshold boundary.Classification remains applicable when the excursion-set boundary does not coincide with the output's threshold level set.Source: §2.2 · p. 3
- The authors motivate classification by the limited need to reconstruct the full response, its potential use to initialize nearby-threshold boundaries, and the expense and dimensional scaling of discontinuous-function approximation.They use SVMs and propose adaptive sampling that favors tangential boundary exploration over uncertainty sampling.Source: §2.2 · pp. 3–4
3. Active learning for excursion set estimation
This section reviews regression and classification approaches to excursion-set estimation, then motivates SVM-based boundary representation and active learning. It contrasts common acquisition strategies and their limitations in refining the boundary.
3. Active learning for excursion set estimation
- SVMs provide an explicit decision function for the boundary and are presented as effective in low-data regimes; the section reviews their properties and established active-learning strategies.The review treats regression-based level-set estimation first, followed by classification-based methods.Source: §3 · p. 4
3.1. Gaussian process regression
- Gaussian-process regression conditions a prior response model on simulator measurements to estimate the response and its uncertainty.The conditional mean is the regressor, and the conditional variance supplies uncertainty estimates.Source: §3.1 · p. 4
3.2. Gaussian process-based active learning
- GP active-learning methods use uncertainty or confidence regions to guide queries near an estimable level set, but depend on regression fidelity and can be unsuitable when outputs are discontinuous.The section describes level-set estimation, Straddle, and maximum-variance acquisition; batch uncertainty is updated while the posterior mean is held fixed.Source: §3.2 · pp. 4–5
3.3. Support vector machines
- The SVM classifier labels simulator samples and uses the sign of its decision function to separate approximate regimes; sparse support vectors and differentiable kernels make function and gradient evaluations tractable.The cited comparison states that kernel SVM training is generally cheaper than GP training in the low-data regime.Source: §3.3 · pp. 5–6
3.4. Active learning for support vector machines
- Common SVM active-learning methods query near the boundary, but uncertainty sampling can repeatedly sample the same region, while diversity-based alternatives face local-refinement or higher-dimensional drawbacks.The section also reviews clustering, which can plateau and deteriorate at moderate dimension, and uncertainty-diversity sampling, whose discrete search may degrade in higher dimensions.Source: §3.4 · pp. 6–7
4. Our proposed approach
The approach represents the estimated regime boundary with an SVM and uses gradient-based sampling to explore it, followed by space-filling batch selection. HMC momentum supports coherent movement along the boundary, while inverse temperature controls sample concentration normal to it.
4. Our proposed approach
- The SVM decision function induces a potential minimized at the estimated boundary, whose analytical gradient makes gradient-based candidate generation inexpensive relative to model evaluations.A greedy space-filling selection discourages repeated queries in the same local boundary region.Source: §4 · pp. 7–8
- The active-learning loop trains an SVM, generates boundary candidates with HMC, and selects a diverse space-filling batch for simulator queries.The process starts from a coarse dataset and repeats until the computational budget is exhausted.Source: §4 · p. 8
4.1. Hamiltonian Monte Carlo
- HMC adds momentum for long, coherent proposals, while a constant mass matrix keeps leapfrog updates explicit and inexpensive.Chains are initialized near the estimated boundary, and parallel shorter chains are used to promote coverage and diversity.Source: §4.1 · p. 9
4.2. Boundary exploration mechanism
- Tangential random-walk exploration becomes difficult in higher dimensions, while ULA’s strong-convexity conditions are not met for Gaussian-kernel SVMs.Constrained MCMC for higher-dimensional tangential exploration is described as essentially equivalent in spread and cost to HMC with large β.Source: §4.2 · p. 10
- HMC’s force acts in the boundary-normal direction, leaving tangential momentum unpenalized; increasing β narrows normal spread without directly changing tangential motion.The target concentrates mass where |f_t(µ)| ≲ 1/√β.Source: §4.2 · pp. 9–10, Figure 1 · p. 11
4.3. Implementation details and computational cost
- About one minute of sequential sampling wall-clock time was reported for a full run, falling to about 10 seconds with five processors.For model evaluations lasting tens of seconds or minutes, the additional CPU time is negligible.Source: §4.3 · p. 11
5. Numerical examples
The numerical examples compare active-learning strategies on analytical boundaries and discontinuous engineering and biological models, using F1 and, where available, Hausdorff distance. Across these settings, HMC with space-filling selection performs especially well as boundary complexity or parameter dimension increases, while classification methods outperform regression on discontinuous outputs.
5.1. Analytical two-dimensional functions
- On increasingly curved analytical boundaries, space-filling methods improve boundary coverage; HMC leads on Hausdorff distance for the more complex cases, while F1 alone can obscure missed regions.For (45), HMC and margin with diversity provide the best performance, and HMC improves Hausdorff distance throughout the iterations.Source: §5.1 · pp. 13, 15, 17, Figure 6 · p. 15, Figure 9 · p. 17
5.2. Dimension scaling
- As dimension rises, HMC’s advantage in Hausdorff distance grows through n = 4, while from n = 4 margin with diversity becomes unstable and HMC is generally more reliable than MALA.The authors report this trend at a fixed sample size of 2,000 points.Source: §5.2 · p. 19, Figure 12 · p. 19, Figure 13 · p. 20
5.3. Rolling ball on a Gaussian terrain
- For the discontinuous rolling-ball output, classification methods reach high F1 scores more efficiently than regression, and Straddle never reaches F1 ≥0.97.At the low dimensionality tested, space-filling classification methods are broadly equivalent, with HMC and margin leading for most targets.Source: §5.3 · p. 21, Figure 14 · p. 21, Figure 15 · p. 21
5.4. Genetic toggle switch
- On the discontinuous genetic toggle-switch boundary, classification outperforms regression, and the proposed approach is the most robust and efficient classification method in the tested budget.It is the only algorithm to reach 0.99 F1 on a non-negligible percentage of seeds.Source: §5.4 · p. 22, Figure 16 · p. 22, Figure 17 · p. 22
5.5. Dam-break shallow water equation
- For the dam-break flow transition, regression struggles with steep output changes, while HMC remains the most sample-efficient method and generally shows a more robust trend.Margin with diversity yields oscillatory scores, and MALA suffers from diffusive tangential motion.Source: §5.5 · p. 24, Figure 18 · p. 24, Figure 19 · p. 24
6. Conclusions
The proposed SVM-based approach uses gradient-driven dynamics to explore regime boundaries tangentially. Across examples, it often improves sample efficiency and robustness, though very high-dimensional costs and discovering additional components remain limitations.
6. Conclusions
- The method uses support vector machines and Hamiltonian Monte Carlo to explore the estimated boundary tangentially, addressing mode collapse in uncertainty-based sampling.Source: §6 · p. 24
- Compact SVM representations keep function and gradient evaluations cheap, making added overhead near-negligible relative to typical model-evaluation costs for complex models.Source: §6 · p. 25
- Across most tested discontinuous simulation examples, the proposed algorithm is most robust to initialization and most likely to reach a target F1 score under a fixed budget.Source: §6 · p. 25
- In very high dimensions, support-vector and gradient costs may grow too large, and refinement targets a connected component already identified by the initial static design.Source: §6 · p. 25