pith. sign in

arxiv: 2606.05247 · v1 · pith:YJ27QE5Dnew · submitted 2026-06-03 · 💻 cs.LG · stat.ML

DiffSlack: Learning under Nonlinear Inequality Constraints via Learnable Slack Variables

Ziqian Wang , Chenxi Fang , Zhen Zhang This is my paper

Pith reviewed 2026-06-28 07:19 UTC · model grok-4.3

classification 💻 cs.LG stat.ML
keywords nonlinear inequality constraintsdifferentiable projectionlearnable slack variablesGauss-Newton methodpath planningconstraint satisfactionneural network layersvehicle trajectory generation
0
0 comments X

The pith

DiffSlack lets neural networks satisfy hundreds of nonlinear inequality constraints by predicting learnable slack variables that warm-start a differentiable projection.

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

The paper introduces DiffSlack to enforce many coupled nonlinear inequalities on neural network outputs without restricting the constraint set or adding heavy overhead. It augments the network prediction with slack variables that convert inequalities to equalities, then applies a damped Gauss-Newton step to project onto the feasible manifold while keeping the whole pipeline differentiable. A two-stage curriculum stabilizes training. On vehicle path planning with 200 constraints from collision avoidance, curvature, and spacing, the method raises planning success rate and geometric satisfaction compared with prior learning baselines at similar inference cost. The hard projection also reduces dependence on the quality of the training labels.

Core claim

DiffSlack reformulates inequalities as equalities with learnable slack variables, which are predicted as part of the augmented network output and provide a data-driven warm start for damped Gauss-Newton projection. The projection layer maps raw predictions onto the augmented feasible manifold while preserving end-to-end differentiability. A two-stage curriculum further stabilizes training and improves constraint satisfaction.

What carries the argument

Learnable slack variables that augment the network output and supply the initial point for the damped Gauss-Newton projection onto the equality-constrained manifold.

If this is right

  • Planning success rate increases while geometric constraint violations decrease on tasks with many coupled nonlinear limits.
  • The hard projection layer makes final outputs less sensitive to imperfections in the supervised training signal.
  • Generated trajectories remain executable in closed-loop simulation and on physical vehicles.
  • Inference cost stays comparable to unconstrained baselines despite the added projection step.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same slack-augmented projection could be inserted into other regression or control networks that must obey geometric or physical limits.
  • If the projection converges for 200 constraints, modest increases in constraint count may remain practical provided the slack predictor improves.
  • Replacing the Gauss-Newton solver with a learned or approximate projection could further reduce per-sample time while retaining differentiability.

Load-bearing premise

The damped Gauss-Newton iteration converges reliably from the learned slack initialization and the curriculum keeps end-to-end training stable.

What would settle it

A test case on the 200-constraint path-planning benchmark where the projection either diverges or produces lower success rates and weaker constraint satisfaction than the strongest learning baseline at matched inference time.

Figures

Figures reproduced from arXiv: 2606.05247 by Chenxi Fang, Zhen Zhang, Ziqian Wang.

Figure 1
Figure 1. Figure 1: Comparison of three approaches to constraint handling in neural networks. (a) Soft constraints incorporate [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall framework and geometric intuition of DiffSlack. The network predicts both the task output [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Experimental environment and G-APF supervision construction. (a) Static path-planning scenario with [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on a representative planning scenario. The upper panels show final trajectories [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Dynamic validation environments. (a) CARLA simulation. (b) Real-world experiment with physical obsta [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Tracking error analysis for three representative real-world runs. Each row shows the lateral error spatial [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ablation study on supervision signal quality under a [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Training dynamics with and without Stage 1 (S1) of the curriculum training strategy, averaged over three [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Ablation study on the maximum iteration count [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Representative failure cases of DiffSlack. Each column shows the raw output (top) and the final output [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
read the original abstract

Enforcing nonlinear inequality constraints in neural networks remains challenging, especially when the output is subject to many coupled constraints. Existing hard constraint methods often impose structural restrictions on the constraint set or introduce substantial computational overhead for large-scale nonlinear problems. Here, we propose DiffSlack, a differentiable projection layer for nonlinear inequality-constrained neural prediction. DiffSlack reformulates inequalities as equalities with learnable slack variables, which are predicted as part of the augmented network output and provide a data-driven warm start for damped Gauss-Newton projection. The projection layer maps raw predictions onto the augmented feasible manifold while preserving end-to-end differentiability. A two-stage curriculum further stabilizes training and improves constraint satisfaction. We evaluate DiffSlack on vehicle path planning with 200 nonlinear inequality constraints from collision avoidance, curvature limits, and waypoint spacing. Compared with existing learning-based baselines, DiffSlack achieves a higher planning success rate and stronger geometric constraint satisfaction under a comparable inference budget. Ablation studies further show that the hard projection layer reduces sensitivity to supervision quality. Closed-loop tracking in CARLA and real-world vehicle experiments confirms the executability of the generated trajectories. These results demonstrate that DiffSlack provides a practical and scalable approach to embedding hard inequality constraints into neural networks for engineering applications.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes DiffSlack, a differentiable projection layer for neural networks subject to many coupled nonlinear inequality constraints. It reformulates the inequalities as equalities using learnable slack variables that are predicted jointly with the network output to warm-start a damped Gauss-Newton solver; the resulting projection is claimed to be end-to-end differentiable. A two-stage curriculum is used to stabilize training. The method is evaluated on a vehicle path-planning task with 200 nonlinear constraints (collision avoidance, curvature, waypoint spacing), where it reports higher planning success rates and better geometric satisfaction than learning-based baselines at comparable inference cost, plus closed-loop validation in CARLA and on a real vehicle.

Significance. If the convergence and differentiability claims hold, DiffSlack would supply a practical, scalable route to hard nonlinear inequality enforcement inside neural predictors without the structural restrictions of many existing projection or barrier methods. The vehicle-planning results and real-world experiments would constitute concrete evidence that the approach can be deployed under realistic constraint counts and inference budgets.

major comments (2)
  1. [§4, §5.2] §4 (Projection Layer) and §5.2 (Experiments): the headline claim of higher success rate and stronger constraint satisfaction rests on the damped Gauss-Newton solver reliably reaching a feasible point from the slack-augmented network output for 200 coupled nonlinear constraints. No iteration counts, convergence-failure rates, or sensitivity plots versus initialization are reported; without these data the performance advantage cannot be attributed to the differentiable hard-projection construction rather than curriculum or post-processing.
  2. [§3.2] §3.2 (Two-stage Curriculum): the assertion that the curriculum “stabilizes end-to-end training while preserving differentiability” is load-bearing for the training procedure, yet the manuscript provides neither the exact schedule of the two stages nor an ablation that isolates its effect on projection convergence versus final constraint satisfaction.
minor comments (2)
  1. [§3] Notation for the slack-augmented output vector and the Jacobian of the equality map should be introduced once and used consistently; several equations reuse the same symbol for different quantities.
  2. [Figure 3] Figure 3 (success-rate bar plot) lacks error bars or statistical significance markers; the numerical values should be stated in the caption or a table.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments below and will revise the manuscript to incorporate additional experimental details and analyses where appropriate.

read point-by-point responses

  1. Referee: [§4, §5.2] §4 (Projection Layer) and §5.2 (Experiments): the headline claim of higher success rate and stronger constraint satisfaction rests on the damped Gauss-Newton solver reliably reaching a feasible point from the slack-augmented network output for 200 coupled nonlinear constraints. No iteration counts, convergence-failure rates, or sensitivity plots versus initialization are reported; without these data the performance advantage cannot be attributed to the differentiable hard-projection construction rather than curriculum or post-processing.

    Authors: We agree that the absence of explicit convergence diagnostics limits the strength of the attribution. In the revised manuscript we will add, in §5.2 and the supplementary material, (i) average and maximum iteration counts per projection call across the test set, (ii) the fraction of projections that failed to reach feasibility within the iteration budget, and (iii) sensitivity plots of final constraint violation versus initialization distance for representative constraint sets. These additions will allow readers to verify that the damped Gauss-Newton solver consistently reaches feasible points from the slack-augmented predictions and that the reported gains are not solely due to the curriculum or post-processing. revision: yes

  2. Referee: [§3.2] §3.2 (Two-stage Curriculum): the assertion that the curriculum “stabilizes end-to-end training while preserving differentiability” is load-bearing for the training procedure, yet the manuscript provides neither the exact schedule of the two stages nor an ablation that isolates its effect on projection convergence versus final constraint satisfaction.

    Authors: We acknowledge that the two-stage curriculum description in §3.2 is insufficiently detailed. In the revision we will (i) state the precise epoch ranges, loss-weight schedules, and projection-activation thresholds used in each stage, and (ii) add an ablation that trains identical networks with and without the curriculum, reporting both projection convergence statistics and final geometric constraint satisfaction. This will isolate the curriculum’s contribution to training stability and constraint satisfaction while preserving the end-to-end differentiability of the projection layer. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain

full rationale

The abstract and description present DiffSlack as a new differentiable projection layer that augments the network output with learnable slack variables to warm-start an existing damped Gauss-Newton solver for nonlinear equalities. No quoted step reduces a claimed prediction or result to its own inputs by construction, renames a known pattern, or relies on a self-citation chain for a uniqueness theorem. The two-stage curriculum is described as a stabilization heuristic, not a definitional loop. Empirical claims rest on external benchmarks (vehicle planning, CARLA, real-world tests) rather than internal fitting. This is the common honest case of a self-contained engineering contribution.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

Based solely on the abstract, the central innovation is the introduction of learnable slack variables; no other free parameters or axioms are identifiable without the full text.

invented entities (1)
  • learnable slack variables no independent evidence
    purpose: To reformulate nonlinear inequalities as equalities and provide warm start for projection
    Introduced as part of the DiffSlack method to enable the projection layer.

pith-pipeline@v0.9.1-grok · 5748 in / 1139 out tokens · 30978 ms · 2026-06-28T07:19:55.810291+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

51 extracted references · 32 canonical work pages · 4 internal anchors

  1. [1]

    Differentiable convex optimization layers, in: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alch´e-Buc, F., Fox, E., Garnett, R

    Agrawal, A., Amos, B., Barratt, S., Boyd, S., Diamond, S., Kolter, J.Z., 2019. Differentiable convex optimization layers, in: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alch´e-Buc, F., Fox, E., Garnett, R. (Eds.), Advances in Neural Information Processing Systems, Curran Associates, Inc

    2019

  2. [2]

    D., Xu, X., Grizzle, J

    Ames, A.D., Xu, X., Grizzle, J.W., Tabuada, P., 2017. Control barrier function based quadratic programs for safety critical systems. IEEE Transactions on Automatic Control 62, 3861–3876. doi:10.1109/TAC.2016.2638961

    work page doi:10.1109/tac.2016.2638961 2017

  3. [3]

    OptNet: Differentiable optimization as a layer in neural networks, in: Proceedings of the International Conference on Machine Learning, pp

    Amos, B., Kolter, J.Z., 2017. OptNet: Differentiable optimization as a layer in neural networks, in: Proceedings of the International Conference on Machine Learning, pp. 136–145

    2017

  4. [4]

    CasADi: A software frame- work for nonlinear optimization and optimal control

    Andersson, J.A.E., Gillis, J., Horn, G., Rawlings, J.B., Diehl, M., 2019. CasADi: A software frame- work for nonlinear optimization and optimal control. Mathematical Programming Computation 11, 1–36. doi:10.1007/s12532-018-0139-4

    work page doi:10.1007/s12532-018-0139-4 2019

  5. [5]

    Curriculum learning, in: Proceedings of the Interna- tional Conference on Machine Learning, pp

    Bengio, Y ., Louradour, J., Collobert, R., Weston, J., 2009. Curriculum learning, in: Proceedings of the Interna- tional Conference on Machine Learning, pp. 41–48

    2009

  6. [6]

    Implicit differentiation for fast hyperparameter selection in non-smooth convex learning

    Bertrand, Q., Klopfenstein, Q., Massias, M., Blondel, M., Vaiter, S., Gramfort, A., Salmon, J., 2022. Implicit differentiation for fast hyperparameter selection in non-smooth convex learning. Journal of Machine Learning Research 23, 1–43

    2022

  7. [7]

    Efficient and modular implicit differentiation, in: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A

    Blondel, M., Berthet, Q., Cuturi, M., Frostig, R., Hoyer, S., Llinares-Lopez, F., Pedregosa, F., Vert, J.P., 2022. Efficient and modular implicit differentiation, in: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (Eds.), Advances in Neural Information Processing Systems, Curran Associates, Inc.. pp. 5230–5242

    2022

  8. [8]

    Convex Optimization

    Boyd, S.P., Vandenberghe, L., 2004. Convex Optimization. Cambridge Univ. Press

    2004

  9. [9]

    Physics-informed neural networks with hard linear equality constraints

    Chen, H., Flores, G.E.C., Li, C., 2024. Physics-informed neural networks with hard linear equality constraints. Computers & Chemical Engineering 189, 108764. doi:10.1016/j.compchemeng.2024.108764

    work page doi:10.1016/j.compchemeng.2024.108764 2024

  10. [10]

    Theory-guided hard constraint projection (HCP): A knowledge-based data-driven scientific machine learning method

    Chen, Y ., Huang, D., Zhang, D., Zeng, J., Wang, N., Zhang, H., Yan, J., 2021. Theory-guided hard constraint projection (HCP): A knowledge-based data-driven scientific machine learning method. Journal of Computational Physics 445, 110624. doi:https://doi.org/10.1016/j.jcp.2021.110624

    work page doi:10.1016/j.jcp.2021.110624 2021

  11. [11]

    End-to-end learning for optimization via constraint-enforcing approximators

    Cristian, R., Harsha, P., Perakis, G., Quanz, B.L., Spantidakis, I., 2023. End-to-end learning for optimization via constraint-enforcing approximators. Proceedings of the AAAI Conference on Artificial Intelligence 37, 7253–

    2023

  12. [12]

    doi:10.1609/aaai.v37i6.25884

    work page doi:10.1609/aaai.v37i6.25884

  13. [13]

    Physics-informed neural particle flow for the bayesian update step

    Csuzdi, D., B ´ecsi, T., T ¨or˝o, O., 2026. Physics-informed neural particle flow for the bayesian update step. Knowledge-Based Systems 346, 116209. doi:https://doi.org/10.1016/j.knosys.2026.116209

    work page doi:10.1016/j.knosys.2026.116209 2026

  14. [14]

    Knowledge-augmented deep learning and its ap- plications: A survey

    Cui, Z., Gao, T., Talamadupula, K., Ji, Q., 2025. Knowledge-augmented deep learning and its ap- plications: A survey. IEEE Transactions on Neural Networks and Learning Systems 36, 2133–2153. doi:10.1109/TNNLS.2023.3338619

    work page doi:10.1109/tnnls.2023.3338619 2025

  15. [15]

    Numerische Mathematik , author =

    Dijkstra, E.W., 1959. A note on two problems in connexion with graphs. Numerische Mathematik 1, 269–271. doi:10.1007/BF01386390. 17 APREPRINT- JUNE5, 2026

    work page doi:10.1007/bf01386390 1959

  16. [16]

    Path planning for autonomous vehicles in unknown semi-structured environments

    Dolgov, D., Thrun, S., Montemerlo, M., Diebel, J., 2010. Path planning for autonomous vehicles in unknown semi-structured environments. International Journal of Robotics Research 29, 485–501. doi:10.1177/0278364909359210

    work page doi:10.1177/0278364909359210 2010

  17. [17]

    DC3: A learning method for optimization with hard constraints, in: Proceedings of the International Conference on Learning Representations

    Donti, P., Rolnick, D., Kolter, J.Z., 2021. DC3: A learning method for optimization with hard constraints, in: Proceedings of the International Conference on Learning Representations

    2021

  18. [18]

    CARLA: An open urban driving simulator, in: Proceedings of the Conference on Robot Learning, pp

    Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., Koltun, V ., 2017. CARLA: An open urban driving simulator, in: Proceedings of the Conference on Robot Learning, pp. 1–16

    2017

  19. [19]

    Curvature measures

    Federer, H., 1959. Curvature measures. Transactions of the American Mathematical Society 93, 418–491

    1959

  20. [20]

    Lagrangian duality for constrained deep learning, in: Joint European conference on machine learning and knowledge discovery in databases, Springer

    Fioretto, F., Van Hentenryck, P., Mak, T.W., Tran, C., Baldo, F., Lombardi, M., 2020. Lagrangian duality for constrained deep learning, in: Joint European conference on machine learning and knowledge discovery in databases, Springer. pp. 118–135

    2020

  21. [21]

    Gammell, J.D., Srinivasa, S.S., Barfoot, T.D., 2014. Informed RRT*: Optimal sampling-based path planning focused via direct sampling of an admissible ellipsoidal heuristic, in: Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 2997–3004

    2014

  22. [22]

    Acda: Anatomically constrained distribution alignment for robust medical image segmentation

    Gao, Y ., Wang, X., Gao, X., 2026. Acda: Anatomically constrained distribution alignment for robust medical image segmentation. Knowledge-Based Systems 341, 115749. doi:10.1016/j.knosys.2026.115749

    work page doi:10.1016/j.knosys.2026.115749 2026

  23. [23]

    On Differentiating Parameterized Argmin and Argmax Problems with Application to Bi-level Optimization

    Gould, S., Fernando, B., Cherian, A., Anderson, P., Santa Cruz, R., Guo, E., 2016. On differentiating parame- terized argmin and argmax problems with application to bi-level optimization. arXiv preprint arXiv:1607.05447 doi:10.48550/arXiv.1607.05447

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1607.05447 2016

  24. [24]

    Pinet: Optimizing hard-constrained neural networks with orthogonal projection layers, in: Proceedings of the International Conference on Learning Representations

    Grontas, P.D., Terpin, A., Balta, E.C., D’Andrea, R., Lygeros, J., 2026. Pinet: Optimizing hard-constrained neural networks with orthogonal projection layers, in: Proceedings of the International Conference on Learning Representations

    2026

  25. [25]

    21, 2022

    Guo, Y ., Guo, Z., Wang, Y ., Yao, D., Li, B., Li, L., 2024. A survey of trajectory planning methods for au- tonomous driving—part i: Unstructured scenarios. IEEE Transactions on Intelligent Vehicles 9, 5407–5434. doi:10.1109/TIV .2023.3337318

    work page doi:10.1109/tiv 2024

  26. [26]

    Physics-informed neural networks with hard nonlinear equality and inequality constraints

    Iftakher, A., Golder, R., Roy, B.N., Faruque Hasan, M., 2026. Physics-informed neural networks with hard nonlinear equality and inequality constraints. Computers & Chemical Engineering 204, 109418. doi:10.1016/j.compchemeng.2025.109418

    work page doi:10.1016/j.compchemeng.2025.109418 2026

  27. [27]

    A survey on projection neural networks and their applications

    Jin, L., Li, S., Hu, B., Liu, M., 2019. A survey on projection neural networks and their applications. Applied Soft Computing 76, 533–544. doi:https://doi.org/10.1016/j.asoc.2019.01.002

    work page doi:10.1016/j.asoc.2019.01.002 2019

  28. [28]

    Real-time obstacle avoidance for manipulators and mobile robots, in: Proceedings

    Khatib, O., 1985. Real-time obstacle avoidance for manipulators and mobile robots, in: Proceedings. 1985 IEEE International Conference on Robotics and Automation, pp. 500–505. doi:10.1109/ROBOT.1985.1087247

    work page doi:10.1109/robot.1985.1087247 1985

  29. [29]

    Kim, M., Kim, H., 2022. Projection-aware deep neural network for dc optimal power flow without constraint violations, in: 2022 IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids (SmartGridComm), pp. 116–121. doi:10.1109/SmartGridComm52983.2022.9961047

    work page doi:10.1109/smartgridcomm52983.2022.9961047 2022

  30. [30]

    ENFORCE: Nonlinear Constrained Learning with Adaptive-depth Neural Projection

    Lastrucci, G., Schweidtmann, A.M., 2025. ENFORCE: Nonlinear constrained learning with adaptive-depth neural projection. arXiv preprint arXiv:2502.06774 doi:10.48550/arXiv.2502.06774

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2502.06774 2025

  31. [31]

    Learning to solve optimization problems with hard linear constraints

    Li, M., Kolouri, S., Mohammadi, J., 2023. Learning to solve optimization problems with hard linear constraints. IEEE Access PP, 1–1. doi:10.1109/ACCESS.2023.3285199

    work page doi:10.1109/access.2023.3285199 2023

  32. [32]

    Liang, E., Chen, M., 2025. Efficient bisection projection to ensure neural-network solution feasibility for op- timization over general set, in: Proceedings of the International Conference on Machine Learning, PMLR. pp. 37071–37099

    2025

  33. [33]

    Homeomorphic projection to ensure neural-network solu- tion feasibility for constrained optimization

    Liang, E., Chen, M., Low, S.H., 2024. Homeomorphic projection to ensure neural-network solu- tion feasibility for constrained optimization. Journal of Machine Learning Research 25, 1–55. URL: http://jmlr.org/papers/v25/23-1577.html

    2024

  34. [34]

    HardNet: Hard-constrained neural networks with universal approximation guarantees

    Min, Y ., Azizan, N., 2024. HardNet: Hard-constrained neural networks with universal approximation guarantees. arXiv preprint arXiv:2410.10807

    arXiv 2024

  35. [35]

    Self-supervised primal-dual learning for constrained optimization

    Park, S., Van Hentenryck, P., 2023. Self-supervised primal-dual learning for constrained optimization. Proceed- ings of the AAAI Conference on Artificial Intelligence 37, 4052–4060. doi:10.1609/aaai.v37i4.25520

    work page doi:10.1609/aaai.v37i4.25520 2023

  36. [36]

    Polack, P., Altch ´e, F., d’Andr´ea Novel, B., de La Fortelle, A., 2017. The kinematic bicycle model: A consis- tent model for planning feasible trajectories for autonomous vehicles?, in: Proceedings of the IEEE Intelligent Vehicles Symposium, Los Angeles, CA, USA. pp. 812–818. 18 APREPRINT- JUNE5, 2026

    2017

  37. [37]

    Raissi, P

    Raissi, M., Perdikaris, P., Karniadakis, G., 2019. Physics-informed neural networks: A deep learning frame- work for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics 378, 686–707. doi:10.1016/j.jcp.2018.10.045

    work page doi:10.1016/j.jcp.2018.10.045 2019

  38. [38]

    Model Predictive Control: Theory, Computation, and Design

    Rawlings, J., Mayne, D., Diehl, M., 2017. Model Predictive Control: Theory, Computation, and Design. Nob Hill Publishing

    2017

  39. [39]

    PythonRobotics: a Python code collection of robotics algorithms

    Sakai, A., Ingram, D., Dinius, J., Chawla, K., Raffin, A., Paques, A., 2018. PythonRobotics: a Python code collection of robotics algorithms. arXiv preprint arXiv:1808.10703 doi:10.48550/arXiv.1808.10703

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1808.10703 2018

  40. [40]

    Unify: A unified policy designing framework for solving integrated constrained optimization and machine learning problems

    Silvestri, M., De Filippo, A., Lombardi, M., Milano, M., 2024. Unify: A unified policy designing framework for solving integrated constrained optimization and machine learning problems. Knowledge-Based Systems 303, 112383. doi:10.1016/j.knosys.2024.112383

    work page doi:10.1016/j.knosys.2024.112383 2024

  41. [41]

    RAYEN: Imposition of Hard Convex Constraints on Neural Networks

    Tordesillas, J., Klemm, V ., How, J.P., Hutter, M., 2026. RAYEN: Imposition of hard convex constraints on neural networks. arXiv preprint arXiv:2307.08336 doi:10.48550/arXiv.2307.08336

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2307.08336 2026

  42. [42]

    Constraint guided gradient descent: Training with inequal- ity constraints with applications in regression and semantic segmentation

    Van Baelen, Q., Karsmakers, P., 2023. Constraint guided gradient descent: Training with inequal- ity constraints with applications in regression and semantic segmentation. Neurocomputing 556, 126636. doi:https://doi.org/10.1016/j.neucom.2023.126636

    work page doi:10.1016/j.neucom.2023.126636 2023

  43. [43]

    A predictive safety filter for learning-based control of constrained nonlinear dynamical systems

    Wabersich, K.P., Zeilinger, M.N., 2021. A predictive safety filter for learning-based control of constrained nonlinear dynamical systems. Automatica 129, 109597. doi:https://doi.org/10.1016/j.automatica.2021.109597

    work page doi:10.1016/j.automatica.2021.109597 2021

  44. [44]

    2024 , note =

    Wabersich, K.P., Zeilinger, M.N., 2023. Predictive control barrier functions: Enhanced safety mechanisms for learning-based control. IEEE Transactions on Automatic Control 68, 2638–2651. doi:10.1109/TAC.2022.3175628

    work page doi:10.1109/tac.2022.3175628 2023

  45. [45]

    Mathematical Programming 106(1): 25--57

    W ¨achter, A., Biegler, L.T., 2006. On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear programming. Mathematical Programming 106, 25–57. doi:10.1007/s10107-004-0559-y

    work page doi:10.1007/s10107-004-0559-y 2006

  46. [46]

    LinSATNet: The positive linear satisfiability neural networks, in: Proceedings of the International Conference on Machine Learning, PMLR

    Wang, R., Zhang, Y ., Guo, Z., Chen, T., Yang, X., Yan, J., 2023. LinSATNet: The positive linear satisfiability neural networks, in: Proceedings of the International Conference on Machine Learning, PMLR. pp. 36605– 36625

    2023

  47. [47]

    Barriernet: Differentiable control barrier functions for learning of safe robot control

    Xiao, W., Wang, T.H., Hasani, R., Chahine, M., Amini, A., Li, X., Rus, D., 2023. Barriernet: Differentiable control barrier functions for learning of safe robot control. IEEE Transactions on Robotics 39, 2289–2307. doi:10.1109/TRO.2023.3249564

    work page doi:10.1109/tro.2023.3249564 2023

  48. [48]

    Knowledge-informed generative adversarial network for functional calibration of computer models

    Yu, Y ., Atamturktur, S., 2023. Knowledge-informed generative adversarial network for functional calibration of computer models. Knowledge-Based Systems 263, 110294. doi:https://doi.org/10.1016/j.knosys.2023.110294

    work page doi:10.1016/j.knosys.2023.110294 2023

  49. [49]

    Zeng, H., Yang, C., Zhou, Y ., Yang, C., Guo, Q., 2024. Glinsat: The general linear satisfiability neural network layer by accelerated gradient descent, in: Globerson, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J., Zhang, C. (Eds.), Advances in Neural Information Processing Systems, Curran Associates, Inc.. pp. 122584– 122615. doi:10.5220...

    work page doi:10.52202/079017-3896 2024

  50. [50]

    Control barrier function-guided deep reinforcement learning for decision-making of autonomous vehicle at on-ramp merging

    Zhang, C., Dai, L., Zhang, H., Wang, Z., 2025. Control barrier function-guided deep reinforcement learning for decision-making of autonomous vehicle at on-ramp merging. IEEE Transactions on Intelligent Transportation Systems 26, 8919–8932. doi:10.1109/TITS.2025.3540862

    work page doi:10.1109/tits.2025.3540862 2025

  51. [51]

    Training neural net- works with classification rules for incorporating domain knowledge

    Zhang, W., Liu, F., Nguyen, C.M., Ou Yang, Z.L., Ramasamy, S., Foo, C.S., 2024. Training neural net- works with classification rules for incorporating domain knowledge. Knowledge-Based Systems 294, 111716. doi:10.1016/j.knosys.2024.111716. 19

    work page doi:10.1016/j.knosys.2024.111716 2024