OASIS: SVM Active Learning for Set Inversion

• OASIS is a framework that reformulates the set inversion problem as binary classification using a Gaussian RBF kernel SVM.
• It employs an active learning strategy with boundary projection to minimize expensive forward model evaluations and accurately capture complex decision boundaries.
• Empirical studies demonstrate that OASIS achieves high accuracy in high-dimensional settings while offering computational benefits over traditional interval-analysis methods.

OASIS, short for Optimization-based Active learning for Set Inversion with Support Vector Machine, is a framework for solving the set inversion problem by recasting it as a binary classification task over the input space. Given a forward map $f : X \to Y$ and a target set $S \subset Y$, OASIS seeks the inversion set $X_S = \{x \in X : f(x) \in S\}$ by learning a decision boundary that approximates $\partial X_S$, while minimizing expensive evaluations of the forward model. The method combines a Gaussian RBF kernel SVM with active learning, specifically margin-based sampling via boundary projection, and is intended for high-dimensional and computationally expensive nonlinear models (Nguyen et al., 2021).

1. Problem formulation and set inversion viewpoint

Set inversion asks for the preimage of a target region under a forward map. In the formulation used by OASIS, if $f : X \to Y$ is the forward map and $S \subset Y$ is the target set, then the inversion set is

$X_S = \{ x \in X : f(x) \in S \}.$

The central objective is not merely pointwise evaluation of $f$, but recovery of the geometry of the subset of $X$ whose image falls inside $S$ (Nguyen et al., 2021).

OASIS defines a binary label on $S \subset Y$0 by

$S \subset Y$1

Equivalently, the classifier

$S \subset Y$2

encodes membership in the preimage. The set inversion problem is therefore reformulated as the task of learning a decision function on $S \subset Y$3 whose positive region approximates $S \subset Y$4 and whose zero-level set approximates $S \subset Y$5.

This formulation changes the computational emphasis. Rather than constructing interval enclosures or exhaustive subdivisions of the domain, OASIS learns the boundary of the inversion set directly in the original input space. This suggests that the method is particularly suited to settings where evaluations of $S \subset Y$6 are expensive and the principal difficulty lies in locating the boundary efficiently.

2. Classification formulation and SVM model

OASIS trains a support vector machine to approximate $S \subset Y$7 by a decision function $S \subset Y$8. The predicted inversion set is

$S \subset Y$9

and the learned decision boundary

$X_S = \{x \in X : f(x) \in S\}$0

is used as an approximation to $X_S = \{x \in X : f(x) \in S\}$1 (Nguyen et al., 2021).

The framework uses a Gaussian Radial Basis Function kernel

$X_S = \{x \in X : f(x) \in S\}$2

The underlying soft-margin SVM is posed in standard primal form as

$X_S = \{x \in X : f(x) \in S\}$3

subject to

$X_S = \{x \in X : f(x) \in S\}$4

The dual form used in the paper is

$X_S = \{x \in X : f(x) \in S\}$5

subject to

$X_S = \{x \in X : f(x) \in S\}$6

The resulting kernel expansion is

$X_S = \{x \in X : f(x) \in S\}$7

A key theoretical point used by OASIS is the universality of the GRBF kernel. With a proper $X_S = \{x \in X : f(x) \in S\}$8, the GRBF kernel can separate any finite labeled dataset in $X_S = \{x \in X : f(x) \in S\}$9, ensuring zero empirical error when the data are separable. In the OASIS procedure, this result motivates calibration of $\partial X_S$0 so that the current labeled set is fit with zero empirical error whenever possible. The intended effect is a sharp decision margin and an accurate boundary approximation.

3. Active learning mechanism and algorithmic workflow

The distinctive element of OASIS is its active learning strategy. Because obtaining a label requires evaluating the forward model $\partial X_S$1 and testing whether $\partial X_S$2, the expensive part of the computation is data acquisition rather than classification itself. OASIS therefore concentrates sampling near the current decision boundary, where uncertainty is highest (Nguyen et al., 2021).

At iteration $\partial X_S$3, the method draws a random point $\partial X_S$4 in the state space and projects it onto the current boundary by solving

$\partial X_S$5

Sequential Quadratic Programming is used for this projection problem. The resulting point $\partial X_S$6 lies on the estimated boundary and is queried to obtain its true label. This is a margin-sampling strategy expressed as a constrained optimization problem rather than as a discrete uncertainty score.

The workflow is specified as follows. The inputs are a state space $\partial X_S$7, a forward map $\partial X_S$8, a target set $\partial X_S$9, an initial seed size $f : X \to Y$0, a query budget $f : X \to Y$1, and SVM hyperparameters including $f : X \to Y$2 and an initial $f : X \to Y$3. The algorithm then:

1. samples $f : X \to Y$4 points uniformly at random from $f : X \to Y$5;
2. labels them by $f : X \to Y$6;
3. repeatedly trains the kernel SVM on the current dataset;
4. increases $f : X \to Y$7 until empirical error on the current labeled set is minimized, ideally zero;
5. samples a random $f : X \to Y$8;
6. projects $f : X \to Y$9 to the current boundary using SQP;
7. evaluates $S \subset Y$0 and queries the label at the projected point;
8. appends the new labeled point and continues until the budget $S \subset Y$1 is exhausted.

The experiments use a fixed query budget as the stopping criterion. The paper also notes that, in practice, stabilization of the margin, a validation accuracy plateau, or a maximum query cost could be used instead. This suggests that the framework is operationally flexible, even though the reported studies emphasize fixed-budget execution.

4. Computational properties and relation to interval-analysis methods

OASIS is positioned against SIVIA and VISIA, which approach set inversion by interval analysis and recursive subpartitioning. In SIVIA, the exact preimage $S \subset Y$2 is enclosed by inner and outer unions of interval boxes, $S \subset Y$3, using an inclusion function

$S \subset Y$4

VISIA is a vectorized Matlab implementation of SIVIA that returns colored certainty boxes: blue for outside, red for inside, and yellow for uncertain.

OASIS departs from that paradigm in two ways. First, it learns a boundary in the original input space rather than recursively subdividing $S \subset Y$5. Second, its computational burden depends mainly on the number of labeled points and support vectors, not exponentially on the dimension $S \subset Y$6 (Nguyen et al., 2021).

The principal per-iteration costs are SVM training, boundary projection, and forward-model evaluation. Kernel SVM solvers typically incur $S \subset Y$7 memory and between $S \subset Y$8 and $S \subset Y$9 time in the worst case. Boundary projection by SQP requires evaluating $X_S = \{ x \in X : f(x) \in S \}.$0 and its gradient, with each evaluation costing $X_S = \{ x \in X : f(x) \in S \}.$1 for dimension $X_S = \{ x \in X : f(x) \in S \}.$2 and support-vector count $X_S = \{ x \in X : f(x) \in S \}.$3. Labeling cost is dominated by evaluating $X_S = \{ x \in X : f(x) \in S \}.$4 and the membership test. Prediction after training is comparatively cheap because evaluating $X_S = \{ x \in X : f(x) \in S \}.$5 is linear in $X_S = \{ x \in X : f(x) \in S \}.$6.

The method is described as relatively robust to increasing dimension because it avoids space partitioning. At the same time, the paper does not claim immunity to the curse of dimensionality. Kernel evaluations still become more expensive with $X_S = \{ x \in X : f(x) \in S \}.$7, and complex boundaries may require more samples and more support vectors. A common misunderstanding is therefore to treat OASIS as dimension-independent; the paper supports only the narrower claim that its computational cost is relatively robust to increasing dimension compared with interval-subdivision methods.

5. Empirical studies and comparative performance

The paper evaluates OASIS on synthetic geometric inversion problems, high-dimensional spheres, and a Lotka–Volterra predator–prey example, and compares it against VISIA (Nguyen et al., 2021).

For two-dimensional shapes defined through $X_S = \{ x \in X : f(x) \in S \}.$8 and $X_S = \{ x \in X : f(x) \in S \}.$9 on $f$0, OASIS used 100 random seeds plus 400 active queries and achieved average accuracy approximately $f$1 across the circle, ring, and doughnut tasks. VISIA achieved approximately $f$2. OASIS trained more slowly in these low-dimensional cases, but prediction per point was faster.

For high-dimensional spheres,

$f$3

OASIS maintained accuracy above $f$4 from 3D through 8D, while VISIA succeeded in 3D and 4D and failed beyond 4D because of memory exhaustion.

In the Lotka–Volterra case,

$f$5

with $f$6, $f$7, $f$8, $f$9,

$X$0

OASIS used 400 random plus 400 active queries and recovered $X$1 with $X$2 accuracy.

Setting	OASIS	VISIA / VSIVIA
2D average accuracy	$X$3	$X$4
3D sphere	$X$5	$X$6
4D sphere	$X$7	$X$8
5D sphere	$X$9	fails beyond 4D
8D sphere	$S$0	fails beyond 4D
Lotka–Volterra	$S$1	not reported as superior

These results support a nuanced interpretation. In low dimensions, VISIA can train much faster because interval subdivision is cheap and OASIS retrains an SVM after each query. In higher dimensions, however, OASIS remains tractable whereas VISIA suffers from exponential growth due to interval subpartitioning and eventually exhausts memory. A plausible implication is that OASIS is best viewed not as a universal replacement for interval methods, but as a high-dimensional alternative whose advantages become more pronounced when the forward model is expensive and the state dimension is moderate to large.

6. Assumptions, limitations, and proposed extensions

OASIS assumes that $S$2 is compact, that $S$3 is smooth or at least continuous enough to make membership tests well posed, and that labels are correct and effectively deterministic. These assumptions matter because the active learner concentrates aggressively near the evolving boundary, where model misspecification or label noise can have amplified effects (Nguyen et al., 2021).

Several limitations follow directly from that design. Boundary-only sampling can overcommit when labels are noisy or when the current classifier misrepresents $S$4. Complex, highly nonconvex, thin, or disconnected preimages may require many support vectors and careful tuning of $S$5. The projection problem can become challenging when $S$6 itself is highly nonconvex. Training cost also grows with the number of labeled samples because the SVM is retrained repeatedly.

The paper discusses practical mitigations rather than claiming full resolution of these issues. Initial random sampling provides exploration before aggressive boundary targeting begins. Cost-sensitive SVMs or reweighting may help when $S$7 occupies a small volume in $S$8. Batch active updates and warm-start SVM solvers are suggested as ways to reduce retraining overhead. If labels are noisy, more exploration or robust SVM variants may be preferable.

The proposed extension space is broad. The classifier could be replaced by Gaussian process classifiers, kernel logistic regression, random forests, or neural networks. Active learning policies could shift from boundary projection to entropy-based sampling, expected model change, query-by-committee, or hybrid exploration–exploitation strategies. Multi-class formulations could address multiple target sets $S$9, and thresholding of $S \subset Y$00 could be used to emulate inner and outer approximations analogous to $S \subset Y$01 and $S \subset Y$02. The paper also points toward sample-complexity theory under low-noise and margin assumptions as a natural direction for formalizing its empirical label efficiency.

In summary, OASIS defines set inversion as learning the preimage boundary of a target set through actively sampled binary classification. Its SVM-based boundary model, SQP-based margin sampling, and avoidance of recursive space partitioning make it a computationally attractive alternative to interval-analysis methods when forward evaluations are expensive and dimensionality is high. Its strongest contribution lies not in replacing geometric inversion theory, but in reframing inversion as a data-efficient decision-boundary estimation problem.