Optimization with Dynamic Constraint Learning (DCL)

Ezgi Oztekin; Figen Oztoprak; S. Ilker Birbil

REVIEW 4 major objections 5 minor 17 references

Optimization with Dynamic Constraint Learning (DCL)

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Unknown constraints can be handled by learning a fresh local surrogate at every optimization step.

desk verdict The core idea is worth engaging with, but the evidence doesn't back the central claim: DCL's output is undefined and the good-looking results come from post-hoc selection of iterates. read the letter →

arxiv 2607.25719 v1 pith:VG36S35U submitted 2026-07-28 cs.LG math.OC

Ezgi Oztekin , Figen Oztoprak , S. Ilker Birbil This is my paper · ORCID

classification cs.LGmath.OC MSC 90C3090C1168T05

keywords dynamicconstraintlearningsurrogate-basedoptimizationunknownconstraintstrustregionlocalgatedata-drivenmixed-integer

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes Dynamic Constraint Learning (DCL), an iterative scheme for constrained optimization when at least one constraint has no known formula and cannot be queried during the run. Instead of fitting one global surrogate to all labeled data before optimizing, DCL fits a simple local surrogate to the data nearest the current iterate, solves a small subproblem confined to the convex hull of that data, and repeats. The paper argues and demonstrates on three experiments that this local, online approach reaches solution quality comparable to global constraint learning while using simpler model classes and much smaller optimization subproblems. If true, DCL offers a practical alternative for problems where a global surrogate is too complex or inaccurate in the region that optimization actually visits.

What carries the argument

The core mechanism is a per-iteration loop: select the nearest-neighbor data subset D_k, fit polynomial or decision-tree models of increasing complexity until validation error stops improving, define the trust region C_k as the convex hull of D_k, then solve a small constrained subproblem (with a penalty for infeasibility) to produce the next iterate. The convex hull trust region is what keeps the local model in a region the data supports, and the dynamic data/model selection is what lets simple local models achieve the accuracy that would require a complex global model.

What would settle it

Run DCL on a problem where the known true optimum is deliberately placed outside the convex hull of the available data; if DCL's returned solution stays far from the true optimum (or the algorithm terminates early), the convex-hull trust region is the cause. Conversely, a data set whose hull contains the optimum but is sparse in between should still work, isolating the support assumption.

Watch

Extended reading notes

Core claim

DCL's central claim is that the location and timing of surrogate learning matter as much as the surrogate itself. By restricting each learned constraint model to a data-supported neighborhood of the current iterate, the method avoids fitting accurate models over regions the optimizer never visits. The trust region is defined as the convex hull of the local training data, so the optimizer cannot extrapolate beyond the data's support. The paper reports that on a synthetic problem DCL reaches near-feasible points where a global surrogate returns a highly infeasible point, and on two real-world case studies it matches global-surrogate solution quality while solving noticeably smaller subproblems

Load-bearing premise

The true optimum must lie inside the convex hull of the data points used around the iterates; if the data does not cover the region around the optimum, DCL's step constraint prevents it from getting there.

Editorial extensions

If this is right

Optimization over unqueryable constraints no longer requires a separate, expensive global training phase; the surrogate is updated as the optimizer moves.
Simple model classes (linear polynomials, shallow trees) can suffice when data density is adequate near the iterates, shrinking MIP subproblems.
The data-supported trust region gives a natural notion of where the learned constraint is trustworthy, preventing the optimizer from exploiting unsupported regions.
DCL can operate in online settings where new data arrives during optimization, since there is no global model to retrain.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

DCL's reliance on a convex hull trust region means it will fail when the true optimum lies outside the region covered by the data; a testable implication is that performance degrades sharply as data coverage near the optimum thins.
The ensemble of local models generated along the optimization path could itself serve as a global surrogate, a possibility the paper raises but does not test.
Switching the trust region to a Mahalanobis ellipsoid or a depth-based region, as the paper suggests, could reduce subproblem size and improve performance on irregular data, but the trade-off is untested.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: the stress-test note is on the mark. The paper introduces a genuinely new heuristic—dynamically refitting local surrogates to nearest-neighbor data, with a data-hull trust region—and writes about it honestly. But the experiments do not show that Algorithm 1 can reliably return a good solution. The Rosenbrock section explicitly says DCL doesn't report an iterate as final, then plots 'the last 20 iterates with relatively small test error.' That is post-hoc selection. The last actual iterate in Table 1 has true constraint value 18.9, infeasible. The diet table oscillates between 3431 and 3683, with the reported solution equal to the best value and no selection rule given. The concrete comparison is against a 'region of attraction,' not a single output. So the abstract's 'can achieve solution quality comparable to global models' is supported only as 'some iterates look nice,' not as a property of the algorithm.

What's good: the framework is coherent, the local model selection (Algorithm 2) is a reasonable way to balance locality and complexity, and the paper is unusually candid—Section 4 admits the termination rule 'proved difficult to trigger' and lists convergence analysis as open. The diet experiment plausibly reproduces the global result with much smaller MIPs, which is a nice sanity check. Related work is handled fairly.

The soft spots are the ones the authors themselves flag, plus one they don't: the convex-hull trust region means the final point is always inside conv(D). If the true optimum lies outside the sampled region, DCL cannot reach it, and the paper never analyzes how often that matters. Combined with unreliable termination, this makes the 'algorithm' more of a template than a method with a defined output.

Who this is for: people working on constraint learning, predict-then-optimize, and surrogate-based optimization will find the idea worth discussing. It's a good paper to put in front of a reading group, mainly as a case study in how hard it is to validate a heuristic. I'd send it to peer review—the idea is new, the writing is clear, and the limitations are honestly stated—but I'd push for major revision: define an output rule, run repeated trials with error bars, release code, and address the hull-confinement issue.

Recommendation: serious referee, not desk reject, with the expectation that the empirical section needs real work.

Referee Report

4 major / 5 minor

Summary. The paper proposes Dynamic Constraint Learning (DCL), an iterative framework for constrained optimization when one constraint function is unknown and unqueryable. At each outer iteration, DCL selects a nearest-neighbor subset of fixed labeled data, fits a local surrogate (whose degree/complexity is chosen by validation), defines a trust region as the convex hull of that subset, and solves a subproblem that minimizes the known objective subject to known analytic constraints, the learned surrogate, and membership in the convex hull. The authors argue that this local, data-adaptive approach can match the solution quality of offline global constraint learning while using simpler models and smaller subproblems, and they report experiments on a synthetic Rosenbrock problem, a palatable diet problem, and a concrete-strength design problem.

Significance. If the central claim were established, DCL would be a practically attractive alternative to the two-phase estimate-then-optimize paradigm for problems with unknown constraints and fixed data. The paper has clear strengths: it engages with relevant literature (global constraint learning, model-based DFO, E2ECO), reuses external baselines [MWB+25], and in the synthetic example the true constraint is known, allowing direct feasibility assessment. The reported local models are indeed simpler (low-degree polynomials, shallow decision trees) and the MIP subproblems are much smaller. However, the empirical evidence as presented does not substantiate the claim that DCL as an algorithm reliably produces the reported solution quality: the termination criterion is acknowledged to be unreliable, and the experiments rely on experimenter-selected iterates or 'regions of attraction' rather than the output of Algorithm 1. No convergence analysis is provided. These gaps are load-bearing for the central claim.

major comments (4)

[§4 (Termination criteria) and §3.1–§3.3] The central claim — that DCL 'can achieve solution quality comparable to that of global models' — is not supported by the experiments because the algorithm's output is not well-defined. Section 4 admits that the termination rule of Algorithm 1 'proved difficult to trigger' and proposes only future criteria. Yet the experiments do not report the final iterate returned by Algorithm 1. In §3.1, Table 1 shows the last iterate (iter 29) has true constraint value 18.9, i.e., infeasible; the paper instead plots 'the last 20 iterates with relatively small test error' as a 'region of attraction' (Figure 3). In §3.2, Table 4 shows the objective oscillating between 3431 and 3683 with no stated selection rule for the reported best value. In §3.3, the comparison is explicitly against a 'region of attraction' rather than a single solution. Thus the empirical evidence demonstrates the existence of favo
[§2.3, Eq. (1) and (2)] The trust region C_k is defined as the convex hull of the local data subset D_k. Therefore every iterate and the final output lie in conv(D), the convex hull of the entire dataset. If the true optimum lies outside the sampled region — a common situation when optima occur at bounds or in sparse-data regions — DCL cannot reach it. The paper does not analyze how often the data hull covers the relevant optimum, nor does it test any instance where the optimum is near a boundary. This is a structural limitation of the proposed framework that should be explicitly quantified or mitigated (e.g., by alternative trust regions like the Mahalanobis ellipsoid mentioned in §2.3) before the general claim of comparable solution quality can be accepted.
[§4 (Convergence analysis) and §1.2] There is no convergence or termination analysis. Section 4 correctly identifies this as open, but it is not merely a theoretical nicety: the empirical behavior shows oscillating iterates and no monotone progress, so the reader cannot infer that DCL converges to a feasible or optimal point even in the limit. Without a convergence guarantee or at least a rigorous characterization of when the step size triggers termination, the algorithm's status as an optimization method is incomplete. A convergence analysis for a simplified setting (e.g., known noise bounds, fixed model class) would strengthen the paper; at minimum, the experiments should demonstrate that some well-defined output rule produces a solution.
[§3 (Experiments)] All experiments are single runs with no error bars, no multiple random seeds, and no sensitivity analysis. Given that the local data selection, train/test splits, and model selection have stochastic components, the reported iterates and timings could be idiosyncratic. This is especially important for the diet problem (§3.2), where the DCL output exactly matches the global-surrogate solution; the reader cannot tell whether this is typical or a single favorable draw. Providing multiple seeds with ranges (or at least a clear statement that results are deterministic up to the solver) is needed to support the comparative claims.

minor comments (5)

[§2.1, Algorithm 1] The requirement line says 'Initial point x0 ∈ X' but the text states that the initial point is 'chosen as a random point within the convex hull of D.' This selection rule should be made explicit in the algorithm or requirement list.
[§2.3] There is a notation inconsistency: the text defines D_k = {i : (z_i, y_i) is in training or test set ...}, using k as both the outer iteration and a set index, while earlier D_k is a set of data points. Clarify the index notation.
[References] The reference [BYcA26] is given as 'arXiv preprint arXiv:2112.06251, 2026.' The arXiv identifier suggests a 2021 submission; please verify the year and version.
[§3.2, Table 4] The table reports a single DCL trajectory. Consider adding the number of runs or seeds used, and whether the reported objective values are the best over runs or a single run.
[General] The phrase 'region of attraction' is used in both §3.1 and §3.3 but is never formally defined. Since it is central to the evaluation, a precise definition (e.g., set of iterates with test error below a threshold) should be given.

Circularity Check

1 steps flagged · score 4.0 of 10

One local comparison is partly tautological (DCL constraint value at iterates is zero by construction), but the central solution-quality comparisons retain independent content.

self definitional [Section 3.1, paragraph after Table 1 and Figure 4 discussion]
"It is possible to observe that the local DCL model value is close to zero at all iterates. This is expected since these points are minimizers of the local model and the constraint of the problem is active at that point by the construction of the problem. ... The true g(x) value is not very far from the local model predictions of DCL as long as the test error is no larger than 0.15."

In subproblem (1)/(2), the trial point is required to satisfy \hat g_k(x) <= 0, or \hat g_k(x) <= s with s penalized, and the paper states the constraint is active at the minimizer by construction. Therefore \hat g_k(x_k) = 0 is imposed by the optimization, not produced as an independent surrogate prediction. Using this value to claim that the true g(x) is close to the 'local model predictions' and plotting |\hat g_k(x_k) - g(x_k)| as 'local error' reduces to measuring the true constraint violation at the iterate, not the surrogate's approximation error. The local-vs-global error comparison is thus partly tautological.

full rationale

The genuine circularity is localized to the synthetic Rosenbrock analysis: the DCL model column is forced to zero by an active constraint in each subproblem, so Figure 4's local-error comparison is not an independent measure of surrogate accuracy. However, this does not reduce the paper's central claim. The diet-problem result is checked against an external implementation from [MWB+25] and DCL reproduces the same mixed-integer solution; the concrete problem uses external data; and the synthetic problem has a known true constraint, with true g values reported. Section 4 contains important admissions: the termination rule 'proved difficult to trigger' and a principled solution-quality assessment is left open because 'neither the true constraint function nor a ground-truth optimal solution is available.' Those are experimental-validity limitations, not circular derivation. The post-hoc selection of a 'region of attraction' from low test-RMSE iterates also weakens the empirical support but is not a fit-renamed-as-prediction in the strict sense. Self-citations [BYcA26], [MBB+24], and [MWB+25] are future work, application motivation, and baseline comparison, respectively; none is load-bearing. Score 4 reflects a partial tautology in a supporting comparison while the core algorithmic idea and its main solution-quality comparisons retain independent content.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the data-hull confinement and on validation error as a proxy for true feasibility; neither is proven. The hyperparameters controlling local neighborhoods are unspecified, so the experiments are not fully reproducible from the text.

free parameters (5)

Initial sample size N_1^k and growth rule for N_t^k
Controls the local neighborhood at each iterate; experimental data sizes vary (e.g., 50-146 in concrete) but the selection rule is not fully specified.
Accuracy threshold ε_model
Algorithm 1 terminates if local model error exceeds this threshold; value is never reported.
Step tolerance ε_step
Algorithm 1 termination condition; authors report it proved difficult to trigger in experiments.
Maximum degree d_max
Limits model complexity in Algorithm 2; in the concrete study depths up to 20 are considered.
Penalty weight ρ_k for relaxation in subproblem (2)
Used when the learned constraint is infeasible inside the trust region; no value or update rule is given.

assumptions (5)

domain assumption The unknown constraint g is a fixed function and each label y_i is a noisy evaluation of g(z_i); the goal is to satisfy g(x) ≤ 0 for the population.
Section 1 learning setup; without this, local surrogates have no well-defined target.
ad hoc to paper Held-out validation error of the local surrogate is a reliable proxy for true constraint error in the trust region.
Algorithm 2 selects models by test RMSE and Algorithm 1 terminates when the threshold is not met; the paper itself flags in Discussion that small surrogate error can still misclassify feasibility when the constraint is active.
domain assumption The optimum lies within the convex hull of the data used (and hence of the whole dataset).
Subproblem (1) with C_k = conv(D_k) confines all iterates to the data hull (Section 2.3).
domain assumption The objective f and known constraints c are analytically available and cheaply evaluable.
Problem statement in Section 2; only g is learned.
domain assumption Iterates can be initiated at a random point in the convex hull of D; x0 must be data-supported.
Section 2.1: initial point chosen as a random point within the convex hull of D.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Optimization with Dynamic Constraint Learning (DCL)." pith.science (2026). https://pith.science/paper/VG36S35U

@misc{pith2026260725719,
  author       = {Pith},
  title        = {Pith review of: Optimization with Dynamic Constraint Learning (DCL)},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VG36S35U}},
  note         = {Machine review of arXiv:2607.25719}
}

read the original abstract

We propose Dynamic Constraint Learning (DCL), a data-driven framework for constrained optimization when constraint functions are unknown and cannot be queried during optimization. At each iteration, the method learns a local surrogate from nearby data and solves a subproblem within a data-supported trust region. Compared with offline global constraint learning, the approach uses local surrogates that adapt to the data distribution during optimization and can achieve solution quality comparable to that of global models while using simpler local models and smaller optimization subproblems. We demonstrate the performance of DCL on a synthetic test problem and two case studies from the literature.

Figures

Figures reproduced from arXiv: 2607.25719 by the authors.

**Figure 2.** First three DCL iterations for the quadratic constraint. Orange: used local subset; gray: [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Synthetic Rosenbrock example. (a) The logarithmically transformed Rosenbrock surface [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Approximation error comparison of the global model and the local models constructed by [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 1 linked inside Pith

[1]

Mathematical Programming Computation , volume=

PyEPO: a PyTorch-based end-to-end predict-then-optimize library for linear and integer programming , author=. Mathematical Programming Computation , volume=. 2024 , publisher=

2024
[2]

AIChE Journal , volume=

Learning surrogate models for simulation-based optimization , author=. AIChE Journal , volume=. 2014 , publisher=

2014
[3]

2008 , publisher=

Engineering design via surrogate modelling: a practical guide , author=. 2008 , publisher=

2008
[4]

2009 , publisher=

Introduction to derivative-free optimization , author=. 2009 , publisher=

2009
[5]

2009 , publisher=

The elements of statistical learning: Data mining, inference, and prediction , author=. 2009 , publisher=

2009
[6]

Operations Research , volume=

Mixed-integer optimization with constraint learning , author=. Operations Research , volume=. 2025 , publisher=

2025
[7]

European Journal of Operational Research , year=

Optimization with constraint learning: a framework and survey , author=. European Journal of Operational Research , year=
[8]

The Journal of Machine Learning Research , volume=

OMLT: Optimization & machine learning toolkit , author=. The Journal of Machine Learning Research , volume=. 2022 , publisher=

2022

Show all 17 references

[9]

INFORMS Journal on Computing , volume=

JANOS: an integrated predictive and prescriptive modeling framework , author=. INFORMS Journal on Computing , volume=. 2022 , publisher=

2022
[10]

Computational Optimization and Applications , volume=

Modeling design and control problems involving neural network surrogates , author=. Computational Optimization and Applications , volume=. 2022 , publisher=

2022
[11]

SIAM Journal on Optimization , volume=

Global convergence of radial basis function trust region derivative-free algorithms , author=. SIAM Journal on Optimization , volume=. 2011 , publisher=

2011
[12]

Applied mathematical modelling , volume=

On data-driven chance constraint learning for mixed-integer optimization problems , author=. Applied mathematical modelling , volume=. 2023 , publisher=

2023
[13]

Mathematics of Operations Research , volume=

Stochastic optimization with decision-dependent distributions , author=. Mathematics of Operations Research , volume=. 2023 , publisher=

2023
[14]

arXiv preprint arXiv:2103.16378 , year =

Kotary, James and Fioretto, Ferdinando and Van Hentenryck, Pascal and Wilder, Bryan , title =. arXiv preprint arXiv:2103.16378 , year =

arXiv
[15]

INFORMS Journal on Computing , year =

Shi, Chenbo and Emadikhiav, Mohsen and Lozano, Leonardo and Bergman, David , title =. INFORMS Journal on Computing , year =
[16]

İlker and Liao, Zhongxing and Bortfeld, Thomas and den Hertog, Dick and Ajdari, Ali , title =

Maragno, Donato and Buti, Gregory and Birbil, Ş. İlker and Liao, Zhongxing and Bortfeld, Thomas and den Hertog, Dick and Ajdari, Ali , title =. 2024 , publisher =

2024
[17]

arXiv preprint arXiv:2112.06251 , year=

Learning with subset stacking , author=. arXiv preprint arXiv:2112.06251 , year=

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Mathematical Programming Computation , volume=

PyEPO: a PyTorch-based end-to-end predict-then-optimize library for linear and integer programming , author=. Mathematical Programming Computation , volume=. 2024 , publisher=

2024

[2] [2]

AIChE Journal , volume=

Learning surrogate models for simulation-based optimization , author=. AIChE Journal , volume=. 2014 , publisher=

2014

[3] [3]

2008 , publisher=

Engineering design via surrogate modelling: a practical guide , author=. 2008 , publisher=

2008

[4] [4]

2009 , publisher=

Introduction to derivative-free optimization , author=. 2009 , publisher=

2009

[5] [5]

2009 , publisher=

The elements of statistical learning: Data mining, inference, and prediction , author=. 2009 , publisher=

2009

[6] [6]

Operations Research , volume=

Mixed-integer optimization with constraint learning , author=. Operations Research , volume=. 2025 , publisher=

2025

[7] [7]

European Journal of Operational Research , year=

Optimization with constraint learning: a framework and survey , author=. European Journal of Operational Research , year=

[8] [8]

The Journal of Machine Learning Research , volume=

OMLT: Optimization & machine learning toolkit , author=. The Journal of Machine Learning Research , volume=. 2022 , publisher=

2022