acceptodds
Under review as a conference paper at ICLR 2027

Optimality-Aware Projection: Aligning Projection Geometry with Constrained Optimization

Abstract

In learning-based constrained optimization pipelines, feasibility-enforcement layers are widely adopted to guarantee strict constraint satisfaction. However, optimizing an unconstrained loss through such feasibility mappings creates a non-equivalent surrogate formulation for the target constrained optimization problem. Specifically, methods based on orthogonal projection exhibit boundary-collapse behavior, which suppresses gradient components normal to active constraints. Consequently, stationary points of this surrogate formulation misalign with the optimality conditions of the target problem, driving convergence to suboptimal solutions. To address this issue, we propose Optimality-Aware (OA) projection , a novel learning mechanism that endows projection-integrated architectures with genuine optimality-seeking capabilities. Without modifying the orthogonal projection layer itself, OA projection injects a KKT-residual-guided geometric correction into the training gradient flow, making it readily embedded into learning-based optimization pipelines. By recovering informative gradients that propel optimization progress along boundaries, this correction favorably reshapes the parameter trajectory and systematically avoids orthogonal-projection-induced spurious stationary points. Theoretically, we establish the differentiability of OA projection, prove its convergence under stochastic gradient updates, and show that under suitable regularity conditions, the converged iterates satisfy the KKT conditions of the target problem. Experimental results on diverse benchmarks, including real-world tasks, confirm that OA projection outperforms state-of-the-art methods by substantially improving the solution quality of hard-constrained learning models while strictly preserving feasibility.

Primary area:
optimization
Keywords:
constrained optimization, end-to-end decision making, differentiable projection layers, hard constraint enforcement, KKT stationarity, optimality-aware projection

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_qJPWOpenReview ↗
Rating
3
Confidence
2(+1)
Informativeness
3
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper studies a limitation of learning through orthogonal projection: although projected outputs satisfy the constraints, projection can remove gradient directions needed to leave a suboptimal boundary face. The proposed Optimality-Aware (OA) projection mixes orthogonal projection with an interior soft-radial map during training. A lagged KKT residual controls the mixing strength, and a decaying schedule eventually recovers orthogonal projection. The paper establishes convergence to surrogate stationary points and conditional KKT consistency results, and evaluates the approach on quadratic and nonconvex programs, trajectory optimization, DC optimal power flow, and graph flow.

Critical strengths
  • A well-motivated distinction between feasibility and optimality. The geometric explanation and the two-dimensional example in Appendix A.1 make the failure mechanism concrete. The paper identifies a meaningful limitation of a commonly used architectural component.
  • A principled, feasibility-preserving construction. The convex combination of two feasible maps preserves feasibility, while the lagged residual gives a clear mechanism for activating the correction. The analysis distinguishes convergence to projection-surrogate stationarity from the stronger, conditional KKT conclusions.
  • Consistent evidence across several optimization settings. The appendix reports multiple seeds, objective gaps, constraint violations, and natural residuals. For example, the mean minimum-cost-flow gap improves from 0.2260 for OP to 0.0810 for OA (Table 10), with very small reported violations. These are useful controls beyond objective value alone.
Critical weaknesses
  1. The practical scope of the KKT guarantee needs clearer qualification. Theorem 5 establishes that cluster points are stationary for the OP-composed objective; it does not itself rule out non-KKT limits. Theorems 6-7 additionally assume convergence to a point and particular face geometry. Theorem 7 also requires the augmented feature matrix to have full row rank, which implies that the feature dimension plus one must be at least the number of training samples. In the DC-OPF experiment, however, Appendix D.3.1 specifies 4,000 fit samples and a width-128 feature backbone, so this condition cannot hold for the described head. That experiment also uses Adam rather than the analyzed SGD updates. The empirical improvements remain valuable, but the paper should distinguish the settings covered by the guarantee from those supported experimentally.

  2. The experiments do not isolate the benefit of KKT feedback from a simpler scheduled mixture. Comparisons against OP and SR show that the full method is useful, but do not establish whether residual regulation is necessary. A fixed mixture and a decaying mixture without residual feedback would be informative controls using the existing benchmarks. Furthermore, the per-sample residual in Eq. (12) requires an additional constrained least-squares calculation. Results under equal iteration budgets do not establish the practical trade-off between solution quality and this training cost; a representative comparison at equal wall-clock time would strengthen the evidence.

  3. The description of the operator is occasionally confusing. The abstract emphasizes a correction to backward gradient flow without modifying the OP layer, whereas Eq. (9) and Algorithm 1 also change the training-time forward output through a convex mixture. Please state this distinction explicitly and clarify that standard OP is used at deployment, as described in Appendix D.3.1.

Questions for authors
  1. Can you explain how Theorem 7 relates to the 4,000-sample, width-128 DC-OPF setup, and identify which guarantees apply to the actual Adam training protocol?
  2. On a representative existing benchmark, how does OA compare with a fixed OP/SR mixture and a decaying mixture without KKT feedback, both at equal iterations and equal wall-clock time? What fraction of training time is spent computing the residual?
  3. Is the forward mixture in Eq. (9) used in every experiment, and is it replaced by OP for validation and deployment in all three application benchmarks?
Reviewer_3EeGOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Optimality-Aware (OA) projection learning method to improve models’ optimality-seeking capabilities. Existing learning-based constrained optimization pipelines usually use a feasibility-enforcement layer, which can suffer from boundary-collapse behavior. To address this challenge, OA projection adds a KKT-residual correction term to the training objective. Experimental results show that OA can improve the solution quality while preserving feasibility in many cases.

Critical strengths
  1. Learning to optimize is a trending research direction, and this work directly aims to improve it.
  2. The proposed OA projection improved solution quality while preserving feasibility.
Critical weaknesses
  1. The paper should improve its presentation. The paper never defines the acronym KKT, which is used throughout.
  2. The new optimization requires an additional KKT residual calculation. The paper lacks a computational cost or time-efficiency analysis to fully understand the value of the OA projection.
Questions for authors

Could you please elaborate on what settings this OA projection optimization method might struggle with?

Reviewer_NKzJOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Ensuring NN solution feasibility with respected to pre-specified constraints is a fundamental issue in ML4CO and ML for real-world decision making. It has attracted attentions in recent years, particularly since the works employing neural network to directly solve optimal power flow (OPF) problems in 2019. The paper focuses on a particular line of feasibility-ensuring approaches, namely employing an orthogonal projection layer in NN. For such an approach, it is known in the community that backpropagation suffers from projection-induced gradient vanishing: the backward signal through an orthogonal projection deletes the gradient components perpendicular to the active constraint wall, because the projection's output does not change when the infeasible input is perturbed along the wall's normal direction (other projection schemes, e.g., gauge mappings and homeomorphic projections, do not suffer from this). Consequently, training can settle at points that look finished but fail the standard optimality certificate (KKT conditions). To fix this, the paper borrows the soft-radial interior map from a recent arXiv paper (Schneider & Kuhn, also cited here) and blends it with the standard projection only during training, with the blend scaled by a KKT-residual. The paper shows that the modified training still converges to a destination set, which still contains the spurious stationary points. Under several assumptions that need to be further verified, including that the training will converge to a single point instead of potential limit cycles, the converged point satisfies the KKT conditions of the original optimization problems (i.e., no spurious stationary points due to the projection-induced gradient vanishing issue). Simulation results, including those for DC-OPF for IEEE test cases demonstrate the optimality improvement performance as compared to existing schemes (but missing some well positioned MLOPF and ML4CO schemes).

Critical strengths
  1. The problem studied is essential to orthogonal-projection based methods for ensuring NN solution feasibility with small optimality loss (at test time).

  2. The convergence analysis (Prop. 12 and Thm. 6) is interesting. The key step in Prop. 12 that one fixed escape direction, the radial direction or the gradient direction, works across all branches of the interior map's piecewise Jacobian is interesting. Intuitively, moving the infeasible point along the radial direction only stretches the ray without changing its direction, so the exact intersection point with any active constraint face remains unchanged. This geometric fact implies that all face-dependent terms in the projection Jacobian must cancel when acting on the radial direction; otherwise, the intersection point would move. For the hard radial projection, the net radial derivative of the projection map is exactly zero; for the smoothed radial projection, the same cancellation holds, leaving only a purely distance-dependent radial term that is identical across all active faces.

  3. Simulation results show the mechanism is working.

Critical weaknesses
  1. The gradient annealing issue is already known in the community, and it has been pointed out by some earlier papers including

Lin, Hung, Yang, Hsieh & Liu, “Escaping from Zero Gradient: Revisiting Action-Constrained Reinforcement Learning via Frank-Wolfe Policy Optimization”, UAI 2021

Brahmanage, Ling & Kumar, “FlowPG: Action-Constrained Policy Gradient with Normalizing Flows”, NeurIPS 2023

and more recently Schneider & Kuhn, “Soft-Radial Projection for Constrained End-to-End Learning”, arXiv 2602.03461 (cited by this paper)

The paper should acknowledge these works and revise the abstract and related statements; as written, readers may incorrectly take the problem's identification as this paper's contribution.

  1. The literature coverage of end-to-end NN schemes for constrained optimization, including the OPF-specific ones, is inadequate. Some early works along the line include

X. Pan, T. Zhao, M. Chen, and S. Zhang, “DeepOPF: A Deep Neural Network Approach for Security-Constrained DC Optimal Power Flow”, IEEE Transactions on Power Systems, 2021.

F. Fioretto, T. Mak and P. V. Hentenryck, "Predicting AC Optimal Power Flows: Combining Deep Learning and Lagrangian Dual Methods”, AAAI 2020.

as well as some ICML/ICLR/NeurIPS papers on ensuring NN solution feasibility. Note that there have been several review papers on the topic thus it should be easy to find a list of relevant works through them. The OPF related-works get parenthetical citations but no discussion. No OPF-specific NN scheme is compared in the simulations, even though the paper's own introduction highlight the DC-OPF performance comparsion.

  1. The forward operator is a convex mix of two existing maps, and the interior map is, as stated in the paper, numerically identical to the baseline they compare against. Other earlier methods, e.g., Gauge mapping based schemes and homeomorphic projection, already avoid the projection-induced gradient vanishing problem. Meanwhile, the KKT residual used in the proposed mechanism is similar to the dual residual, so a comparison to a primal–dual method would be helpful to understand the actual benefit of the proposed mechanism.

  2. The writing is hard to follow. Results are listed without (much) discussion or intuitive explanation, and the theorems are harder to read than they need to be. For instance, Theorem 5 converges to the destination set containing spurious stationary points. It is thus not an optimality statement and the paper should say this plainly. The escape argument's core insight: no single push is large, but infinitely many same-signed vanishing pushes sum to infinity, could have been stated in two sentences for easy understanding.

  3. In simulations, no wall-clock cost is reported. No experiment counts how often converged solutions actually land on spurious points. It was stated that SOTA was compared but the baselines miss some well positioned MLOPF and ML4CO schemes (see above; could also be identified by Google search).

Questions for authors
  1. Under mini-batch training, i.e., Theorem 5's regime, what fraction of converged solutions are actually spurious? Please measure it on the OPF benchmark.

  2. Head-to-head comparison in simulations with a primal–dual method driving the same residual via learned multipliers.

  3. Compare the performance of the proposed mechanism to well-positioned ML4OPF and ML4CO schemes.

  4. Update the literature review and discussions; see weakness #1 and #2.

  5. Clarify whether Theorem 6's convergence assumption holds in general, and whether persistent oscillation within the invariant set (rather than settling at a trap) can occur.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.