Pith. sign in

REVIEW 5 major objections 4 minor 40 references

When GNNs meet symmetry in ILPs: an orbit-based feature augmentation approach

T0 review · 5 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A permutation-equivariant GNN must output identical values for variables in the same orbit of an ILP's symmetry group, so it cannot always predict an optimal solution; the paper proves this and restores accuracy by sampling a small…

desk verdict A correct but elementary theory result plus a genuinely useful orbit-based augmentation scheme; the empirical claims are plausible but rest on heuristic SCIP labels, so the paper deserves a serious referee but needs a rigorous revision. read the letter →

arxiv 2501.14211 v2 pith:UJNRTR3Y submitted 2025-01-24 cs.LG math.OC

Qian Chen , Lei Li , Qian Li , Jianghua Wu , Akang Wang , Ruoyu Sun , Xiaodong Luo , Tsung-Hui Chang
show 1 more author
Qingjiang Shi
This is my paper · ORCID
classification cs.LGmath.OC MSC 90C1068T0705E18
keywords integerlinearprogramminggraphneuralnetworksformulationsymmetryorbitfeatureaugmentationpermutationequivariancebreakingILPsolutionprediction
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

The paper establishes that a GNN trained to predict optimal solutions of an integer linear program (ILP) is structurally limited whenever the ILP has formulation symmetry: because the network is permutation-equivariant over variable nodes and permutation-invariant over constraint nodes, it must return identical values for all variables in the same orbit of the symmetry group, and any optimal solution that distinguishes those variables lies beyond it. This explains why standard GNN solvers underperform on symmetric ILPs, and the authors propose breaking the tie by augmenting each variable with a small feature sampled per orbit from a discrete set. They formulate three principles that augmented features should satisfy—distinguishability, parsimony, and isomorphic consistency—and show that their orbit-based scheme, and a refined version called Orbit+, satisfy them while keeping the augmented feature space far smaller than positional IDs or random noise. If the claim is right, any permutation-equivariant GNN without augmentation is structurally blind on symmetric ILPs, and the orbit-based augmentation is a simple symmetry-aware remedy that improves both accuracy and training speed.

What carries the argument

The central object is the orbit of the ILP's formulation symmetry group, combined with a per-orbit discrete uniform sampling step. For a bipartite variable–constraint representation of an ILP, the symmetry group G partitions variable indices into orbits; the augmentation assigns each variable in an orbit O a distinct sample from {1,...,|O|} without replacement, leaves singleton orbits at zero, and in the Orbit+ variant copies one sampled vector across orbits that the symmetry group moves together. This small random fingerprint breaks the tie identified in Proposition 1 by making the network's input no longer invariant under the symmetry, so the GNN can output different values for formerly symmetric variables. The three principles of distinguishability, augmentation parsimony, and isomorphic consistency are the criteria that make the fingerprint effective rather than merely expressive.

What would settle it

Enumerate all optimal solutions of a small symmetric ILP with a nontrivial orbit; for every pair of isomorphic instances and augmented features that match under a symmetry, check whether the label-adjustment procedure can choose optimal solutions so that the labels also match under that symmetry. If any pair cannot be relabeled, the minimal loss is provably positive on that data, and the paper's assumption that isomorphic consistency is always achievable fails.

Watch

Extended reading notes

Core claim

The central claim is that permutation equivariance over variables and permutation invariance over constraints—properties every standard GNN layer satisfies—combine with formulation symmetry to force equal predictions on symmetric variables (Proposition 1); consequently no such network can always output an optimal solution for a symmetric ILP (Corollary 1). The remedy is to augment the bipartite input: detect the formulation symmetry group, partition variables into orbits, and assign each variable in a nontrivial orbit a distinct integer sampled without replacement from {1,...,|O|}, reusing one sampled vector across orbits that must move together in the Orbit+ variant. The paper proves that training labels that violate the resulting isomorphic-consistency condition force a positive minimal loss (Proposition 2), and it reports that Orbit and Orbit+ reduce Top-m% errors and speed up convergence on bin packing, balanced item placement, and steel mill slab design relative to no augmentation, uniform noise, and positional IDs.

Load-bearing premise

The load-bearing premise is that the SymILO label-adjustment framework can always replace conflicting labels with alternative optimal solutions so that isomorphic consistency holds; if some training instance has no alternative optimum satisfying the consistency condition, the training set contains conflicts and Proposition 2 only guarantees a positive minimal loss, so the reported accuracy gain could be an artifact of label adjustment rather than of the orbit features.

Editorial extensions

If this is right

  • Under the paper's assumptions, any permutation-equivariant GNN with no augmented features produces identical predictions for variables in the same orbit, so on symmetric ILPs it cannot in general predict an optimal solution and some form of symmetry-aware augmentation is required.
  • The orbit-based feature space has cardinality ∏_k |O_k|!, far smaller than the n! space of positional IDs and infinitely smaller than continuous random noise, which the paper identifies as the reason its methods train faster and generalize better.
  • Augmented features that violate isomorphic consistency force the minimal achievable training loss to be positive, so label-adjustment procedures such as SymILO are needed to keep orbit-based training data conflict-free.
  • On the BPP, BIP, and SMSP benchmarks, Orbit and Orbit+ achieve lower Top-m% errors and faster validation-loss convergence than No-Aug, Uniform, and Position, with Orbit+ generally best.
  • Warm-starting the CPLEX solver with Orbit+ predictions yields better objective values at fixed time limits on BIP and SMSP than CPLEX alone, and the predictions show fewer constraint violations than the baselines.

Reading between the lines

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

  • The equal-output obstruction is not specific to GNNs: any predictor that is permutation-equivariant over variables and permutation-invariant over constraints on the same bipartite encoding will suffer the same tie, so orbit-based augmentation is a plausible generic remedy for other equivariant architectures on symmetric combinatorial optimization problems.
  • The paper applies the label-adjustment framework uniformly to every method but reports no ablation that removes it; an experiment training Orbit features with and without label adjustment would separate the contribution of the augmentation itself from the contribution of resolving label conflicts.
  • The same principles could be transferred to other learned solver components, such as branching or node-selection policies, where symmetric variables also create degenerate ties; the paper does not claim this extension.
  • Because the method assumes exact formulation symmetry and needs the detected orbits, its behavior on approximately symmetric or misspecified real-world instances remains untested; the stated scope is ILPs with formulation symmetry.
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.

Referee Report

5 major / 4 minor

Summary. The paper studies the interaction between ILP formulation symmetry and permutation equivariance/invariance of GNNs. It proves (Proposition 1, Corollary 1) that under these standard properties a GNN cannot assign different output values to variables in the same orbit of the ILP symmetry group, so it cannot always predict an optimal solution. To remedy this, the authors propose three principles for constructing augmented variable features (distinguishability, augmentation parsimony, isomorphic consistency), and introduce two orbit-based augmentation schemes, Orbit and Orbit+. The main empirical claim is that these schemes achieve substantially lower Top-m% errors than No-Aug, Uniform, and Position baselines on three benchmark ILP classes (BPP, BIP, SMSP), while also improving training efficiency. The theoretical observation is simple and correct, though the empirical support for the central claim is weakened by uses of non-certified labels, missing error bars, an incompletely specified Orbit+ construction, and an unverified reliance on the SymILO label-adjustment framework.

Significance. If taken at face value, the paper identifies a clean fundamental limitation of equivariant GNNs on symmetric ILPs and offers a practical, parameter-light augmentation scheme; the theoretical result (Proposition 1) is correct under standard assumptions, and the authors provide public source code and reproducible benchmark protocols. The proposed three principles organize the design space of augmentation schemes in a useful way, and the observed qualitative trend that symmetry-aware augmentation improves prediction accuracy is credible. However, the absolute accuracy claims and the specific superiority claims of Orbit+ are not yet fully supported because several load-bearing details of the empirical pipeline and the Orbit+ construction are either unverified or under-specified.

major comments (5)
  1. [Section 5.1 and Section 2.2] The learning task is defined as predicting an optimal solution (Section 2.2), and the isomorphic consistency principle in Section 4.1 is stated for optimal solutions, but the labels used for training and evaluation are only the best solutions found by SCIP within 3,600 seconds and are not certified optimal. The paper never reports optimality gaps or the fraction of instances for which SCIP proved optimality. As a result, the absolute Top-m% errors in Table 2 measure agreement with possibly suboptimal incumbents, not with true optima, and the claimed ability to predict optimal solutions is not supported by the reported numbers. Please report certified optimality statistics, or explicitly reframe the task as predicting approximate incumbents.
  2. [Section 4.2, Orbit+ construction] The paper asserts that the formulation symmetry 'necessitates' that all elements in each column of the orbit matrix O are permuted as an integrated unit, but this property is not true for a general symmetry group. For example, with G = S2 × S2 acting on orbits {1,2} and {3,4}, the permutation that swaps 1 and 2 while fixing 3 and 4 satisfies π(o11) = o12 but maps o21 to itself, not to o22. Moreover, no algorithm is given to compute the column correspondence from the symmetry group. The construction of Orbit+ and the claimed cardinality co+ therefore rest on an additional assumption that must be stated explicitly and verified on the benchmark instances.
  3. [Section 5.3, Eq. (3)] The Top-m% error requires finding π′ = arg min π∈G ∥ˆy − π(y)∥ over the full symmetry group G. For SMSP, Appendix A.8 reports log10|G| ≈ 213, so enumerating G is impossible, yet the paper does not describe how this minimization is computed in practice. Unless this optimization is replaced by an exact or approximate algorithm that is actually used in the experiments, the numbers in Table 2 cannot be reproduced or verified as evaluations of the stated metric.
  4. [Section 5.4, Table 2 and Figure 2] All reported results come from a single random seed, with no standard deviations, confidence intervals, or significance tests. Several differences between methods are small (e.g., BPP Top-70% errors of 1.3 for both Position and Orbit; SMSP Top-90% errors of 50.3 for Orbit+ versus 51.1 for Orbit), so it is unclear whether the claimed improvements are statistically meaningful. Please report results over multiple seeds (or equivalent variance information) and state the number of runs used.
  5. [Section 4.3.1 and Proposition 2] The paper assumes that the SymILO framework can replace conflicting labels with alternative optimal solutions so that the isomorphic consistency principle holds, but it provides no evidence that SymILO actually succeeds on the datasets used. In particular, if the SCIP labels are suboptimal, an alternative optimal solution may not exist for a given instance, and Proposition 2 then implies that a model trained on residual conflicts has nonzero minimal loss. The paper never reports how many training samples remain conflicting after SymILO adjustment, nor does it isolate the effect of SymILO from the effect of the augmentation itself. Please report residual conflict counts or include an ablation without label adjustment to support the claim that orbit features, rather than the label-adjustment machinery, drive the improvement.
minor comments (4)
  1. [Title and Introduction] The title and Section 1 contain stray spacing and a typo in 'indistingushability'; please proofread these passages.
  2. [Section 5.1 and Table 1] The description of SMSP says 'nearly 10,000 constraints', while Table 1 lists 1,000 constraints. These numbers are inconsistent and should be reconciled.
  3. [Appendix A.4] The proof of Proposition 1 begins with an assumption about the loss function ℓ(·) being permutation-invariant with the identity property, but the proof itself never uses this assumption; this introductory sentence appears to be copied from the proof of Proposition 2 and should be removed or moved.
  4. [Section 4.3.1] The explanation of why label-dependent reject sampling is infeasible during testing is condensed; please clarify that at test time the model must accept any drawn augmented feature because no label is available to trigger resampling, whereas during training the labels are known.

Circularity Check

1 steps flagged · score 2.0 of 10

Self-contained equivariance derivation; only minor self-cited SymILO dependency for label adjustment, not a construction-level circularity.

  1. other [Section 4.3.1; Section 5.1]
    "In this paper, we adopt the second approach, as the first one relies on label-dependent reject sampling, which is infeasible during the testing phase when labels are unavailable. Specifically, we utilize the SymILO framework proposed by Chen et al. (2024), which supports dynamically adjusting the labels of the training samples. It jointly optimizes the transformation of solutions and the model parameters, aiming to minimize the prediction error."

    The isomorphic-consistency guarantee needed for Proposition 2's zero-loss regime is not established inside this paper; it is delegated to the authors' own SymILO framework via self-citation. Because the training labels are only SCIP best solutions after 3,600 seconds rather than certified optima, the assumed existence of 'alternative optimal solutions' is not verified, so the absolute claim that orbit features enable prediction of an optimal solution rests on an unverified, self-cited label-adjustment step. This is not a full derivation-equivalent circularity: SymILO is applied uniformly to every augmentation baseline and does not enter the proof of Proposition 1 or the orbit construction, so the relative ranking of Orbit and Orbit+ against Uniform and Position remains independent.

full rationale

The theoretical derivation is self-contained. Proposition 1 follows by composing the assumed permutation equivariance/invariance with the defining property of a formulation symmetry, giving f(A)_i = f(A)_{pi(i)} on each orbit; Corollary 1 is witnessed by an explicit two-optima counterexample. No fitted parameter, learned quantity, or previous result is used there. The orbit-based augmentation scheme is an explicit randomized algorithm, and the cardinality comparison c_{O+} < c_O < c_P < c_U is a direct counting argument over orbit sizes. The isomorphic-consistency principle is defined in Section 4.1 and Proposition 2 is proved from that definition, so it is not imported as a black box. The only self-referential dependency is the use of SymILO in Section 4.3.1 to adjust labels into isomorphic consistency. SymILO is a prior paper by several of the same authors, and the current paper does not prove or verify that its label-replacement assumption holds for the non-certified SCIP labels described in Section 5.1. This is a real empirical-validity caveat and a minor self-citation, but it is not load-bearing for the structural theorems or for the relative comparison of augmentation schemes, since SymILO is applied identically across all methods. The paper's own limitation section also acknowledges that isomorphic consistency applies mainly to supervised tasks with multiple label choices. Overall, no prediction reduces to a fit or to an input by construction; the circularity burden is limited to the self-cited label-adjustment dependency, giving a score of 2.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard equivariance properties of GNNs and on several dataset and pipeline assumptions. No new physical entities or fitted constants are introduced; the only hand-chosen training quantity is the number of augmentation samples per instance. The most consequential assumptions are the treatment of time-limited SCIP incumbents as optimal labels and the unproven availability of orbit-connection structure and SymILO label adjustments.

free parameters (1)
  • Number of augmented samples per training instance = 8
    Chosen by hand as 8; affects stochasticity during training but is not fitted to optimize the reported metric.
assumptions (5)
  • domain assumption GNNs used for ILP prediction are permutation equivariant w.r.t. variable nodes and invariant w.r.t. constraint nodes (Assumption 1).
    This is a property of standard message-passing GNNs; the paper's theory relies on it.
  • domain assumption The ground-truth solution mapping f* is permutation equivariant and invariant (Appendix A.2).
    The isomorphic consistency principle is derived from this, but because an ILP may have multiple optimal solutions, a fixed labeling need not be equivariant; SymILO is invoked to enforce it.
  • standard math The loss function is permutation-invariant and has the identity property (Appendix A.4).
    Used in proof of Proposition 2; MSE and cross-entropy satisfy this.
  • ad hoc to paper The orbit-connection structure used by Orbit+ (all elements of a column are permuted together) holds for the benchmark instances (Section 4.2).
    No general detection algorithm is given; the paper illustrates it with a bin-packing example and applies it without a general procedure.
  • domain assumption Labels obtained from SCIP with a 3600-second time limit are treated as optimal solutions (Section 5.1).
    For SMSP with 23k variables, optimality is not guaranteed; the evaluation target is actually the best incumbent.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When GNNs meet symmetry in ILPs: an orbit-based feature augmentation approach." pith.science (2026). https://pith.science/paper/UJNRTR3Y

@misc{pith2026250114211,
  author       = {Pith},
  title        = {Pith review of: When GNNs meet symmetry in ILPs: an orbit-based feature augmentation approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UJNRTR3Y}},
  note         = {Machine review of arXiv:2501.14211}
}
read the original abstract

A common characteristic in integer linear programs (ILPs) is symmetry, allowing variables to be permuted without altering the underlying problem structure. Recently, GNNs have emerged as a promising approach for solving ILPs. However, a significant challenge arises when applying GNNs to ILPs with symmetry: classic GNN architectures struggle to differentiate between symmetric variables, which limits their predictive accuracy. In this work, we investigate the properties of permutation equivariance and invariance in GNNs, particularly in relation to the inherent symmetry of ILP formulations. We reveal that the interaction between these two factors contributes to the difficulty of distinguishing between symmetric variables. To address this challenge, we explore the potential of feature augmentation and propose several guiding principles for constructing augmented features. Building on these principles, we develop an orbit-based augmentation scheme that first groups symmetric variables and then samples augmented features for each group from a discrete uniform distribution. Empirical results demonstrate that our proposed approach significantly enhances both training efficiency and predictive performance.

Figures

Figures reproduced from arXiv: 2501.14211 by the authors.

Figure 1
Figure 1. Left: An ILP instance where x1 and x2 are symmetric. Middle: A bipartite representation of the ILP instance. The variable and constraint nodes correspond to their counterparts in the ILP instance, with edges connecting them denoting the coefficients of variables in constraints. Right: The outputs for the symmetric variables are identical due to symmetry, thus GNNs cannot correctly predict the optimal solution. 4 [P… view at source ↗
Figure 2
Figure 2. Validation losses of different schemes. 6 CONCLUSION AND LIMITATION In this work, we demonstrated that the interaction between the formulation symmetry of ILPs and the permutation invariance and equivariance properties of GNNs limits the ability of classic GNN architectures to distinguish between symmetric variables. Exploring the potential of feature aug￾mentation to address this limitation, we proposed three guidi… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 33 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Branch and bound methods

    Stephen Boyd and Jacob Mattingley. Branch and bound methods. Notes for EE364b, Stanford University, 2006: 0 07, 2007

    work page 2006

  3. [3]

    Combinatorial optimization and reasoning with graph neural networks

    Quentin Cappart, Didier Ch \'e telat, Elias B Khalil, Andrea Lodi, Christopher Morris, and Petar Veli c kovi \'c . Combinatorial optimization and reasoning with graph neural networks. Journal of Machine Learning Research, 24 0 (130): 0 1--61, 2023

    work page 2023

  4. [4]

    SymILO: A Symmetry-Aware Learning Framework for Integer Linear Optimization

    Qian Chen, Tianjian Zhang, Linxin Yang, Qingyu Han, Akang Wang, Ruoyu Sun, Xiaodong Luo, and Tsung-Hui Chang. Sym ILO : A symmetry-aware learning framework for integer linear optimization. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://arxiv.org/abs/2409.19678

    work page Pith review arXiv 2024

  5. [5]

    On representing mixed-integer linear programs by graph neural networks, 2022

    Ziang Chen, Jialin Liu, Xinshang Wang, Jianfeng Lu, and Wotao Yin. On representing mixed-integer linear programs by graph neural networks, 2022

    work page 2022

  6. [6]

    The scip optimization suite 7.0

    Gerald Gamrath, Daniel Anderson, Ksenia Bestuzheva, Wei-Kun Chen, Leon Eifler, Maxime Gasse, Patrick Gemander, Ambros Gleixner, Leona Gottwald, Katrin Halbig, et al. The scip optimization suite 7.0. 2020

    work page 2020

  7. [7]

    Exact combinatorial optimization with graph convolutional neural networks

    Maxime Gasse, Didier Ch \'e telat, Nicola Ferroni, Laurent Charlin, and Andrea Lodi. Exact combinatorial optimization with graph convolutional neural networks. Advances in neural information processing systems, 32, 2019

    2019

  8. [8]

    The machine learning for combinatorial optimization competition (ml4co): Results and insights

    Maxime Gasse, Simon Bowly, Quentin Cappart, Jonas Charfreitag, Laurent Charlin, Didier Ch \'e telat, Antonia Chmiela, Justin Dumouchelle, Ambros Gleixner, Aleksandr M Kazachkov, et al. The machine learning for combinatorial optimization competition (ml4co): Results and insights. In NeurIPS 2021 competitions and demonstrations track, pp.\ 220--231. PMLR, 2022

    work page 2021

Show all 40 references
  1. [9]

    Miplib 2017: data-driven compilation of the 6th mixed-integer programming library

    Ambros Gleixner, Gregor Hendel, Gerald Gamrath, Tobias Achterberg, Michael Bastubbe, Timo Berthold, Philipp Christophel, Kati Jarck, Thorsten Koch, Jeff Linderoth, et al. Miplib 2017: data-driven compilation of the 6th mixed-integer programming library. Mathematical Programmin...

    2017

  2. [10]

    A gnn-guided predict-and-search framework for mixed-integer linear programming

    Qingyu Han, Linxin Yang, Qian Chen, Xiang Zhou, Dong Zhang, Akang Wang, Ruoyu Sun, and Xiaodong Luo. A gnn-guided predict-and-search framework for mixed-integer linear programming. In The Eleventh International Conference on Learning Representations, 2023

    2023

  3. [11]

    Learning to configure mathematical programming solvers by mathematical programming

    Gabriele Iommazzo, Claudia d’Ambrosio, Antonio Frangioni, and Leo Liberti. Learning to configure mathematical programming solvers by mathematical programming. In Learning and Intelligent Optimization: 14th International Conference, LION 14, Athens, Greece, May 24--28, 2020, Re...

    2020

  4. [12]

    Conflict propagation and component recursion for canonical labeling

    Tommi Junttila and Petteri Kaski. Conflict propagation and component recursion for canonical labeling. In International Conference on Theory and Practice of Algorithms in (Computer) Systems, pp.\ 151--162. Springer, 2011

    2011

  5. [13]

    Symmetry breaking and equivariant neural networks

    S \'e kou-Oumar Kaba and Siamak Ravanbakhsh. Symmetry breaking and equivariant neural networks. arXiv preprint arXiv:2312.09016, 2023

    2023 arXiv

  6. [14]

    Mip-gnn: A data-driven framework for guiding combinatorial solvers

    Elias B Khalil, Christopher Morris, and Andrea Lodi. Mip-gnn: A data-driven framework for guiding combinatorial solvers. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pp.\ 10219--10227, 2022

    2022

  7. [15]

    Improving equivariant networks with probabilistic symmetry breaking

    Hannah Lawrence, Vasco Portilheiro, Yan Zhang, and S \'e kou-Oumar Kaba. Improving equivariant networks with probabilistic symmetry breaking. In ICML 2024 Workshop on Geometry-grounded Representation Learning and Generative Modeling, 2024

    2024

  8. [16]

    Resource allocation optimization based on mixed integer linear programming in the multi-cloudlet environment

    Li Liu and Qi Fan. Resource allocation optimization based on mixed integer linear programming in the multi-cloudlet environment. IEEE Access, 6: 0 24533--24542, 2018

    2018

  9. [17]

    Global optimization method for mixed transportation network design problem: a mixed-integer linear programming approach

    Paramet Luathep, Agachai Sumalee, William HK Lam, Zhi-Chun Li, and Hong K Lo. Global optimization method for mixed transportation network design problem: a mixed-integer linear programming approach. Transportation Research Part B: Methodological, 45 0 (5): 0 808--827, 2011

    2011

  10. [18]

    Pruning by isomorphism in branch-and-cut

    Fran c ois Margot. Pruning by isomorphism in branch-and-cut. Mathematical Programming, 94: 0 71--90, 2002

    2002

  11. [19]

    Exploiting orbits in symmetric ilp

    Fran c ois Margot. Exploiting orbits in symmetric ilp. Mathematical Programming, 98: 0 3--21, 2003

    2003

  12. [20]

    Symmetry in integer linear programming

    Fran c ois Margot. Symmetry in integer linear programming. 50 Years of Integer Programming 1958-2008: From the Early Years to the State-of-the-Art, pp.\ 647--686, 2009

    1958

  13. [21]

    Nauty and traces user’s guide (version 2.5)

    Brendan D McKay and Adolfo Piperno. Nauty and traces user’s guide (version 2.5). Computer Science Department, Australian National University, Canberra, Australia, 2013

    2013

  14. [22]

    Orbit-equivariant graph neural networks

    Matthew Morris, Bernardo Cuenca Grau, and Ian Horrocks. Orbit-equivariant graph neural networks. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=GkJOCga62u

    2024

  15. [23]

    Branch-and-bound algorithms: A survey of recent advances in searching, branching, and pruning

    David R Morrison, Sheldon H Jacobson, Jason J Sauppe, and Edward C Sewell. Branch-and-bound algorithms: A survey of recent advances in searching, branching, and pruning. Discrete Optimization, 19: 0 79--102, 2016

    2016

  16. [24]

    Solving mixed integer programs using neural networks

    Vinod Nair, Sergey Bartunov, Felix Gimeno, Ingrid Von Glehn, Pawel Lichocki, Ivan Lobov, Brendan O'Donoghue, Nicolas Sonnerat, Christian Tjandraatmadja, Pengming Wang, et al. Solving mixed integer programs using neural networks. arXiv preprint arXiv:2012.13349, 2020

    2012 arXiv

  17. [25]

    Constraint orbital branching

    James Ostrowski, Jeff Linderoth, Fabrizio Rossi, and Stefano Smriglio. Constraint orbital branching. In Integer Programming and Combinatorial Optimization: 13th International Conference, IPCO 2008 Bertinoro, Italy, May 26-28, 2008 Proceedings 13, pp.\ 225--239. Springer, 2008

    2008

  18. [26]

    Orbital branching

    James Ostrowski, Jeff Linderoth, Fabrizio Rossi, and Stefano Smriglio. Orbital branching. Mathematical Programming, 126: 0 147--178, 2011

    2011

  19. [27]

    Learning to cut by looking ahead: Cutting plane selection via imitation learning

    Max B Paulus, Giulia Zarpellon, Andreas Krause, Laurent Charlin, and Chris Maddison. Learning to cut by looking ahead: Cutting plane selection via imitation learning. In International conference on machine learning, pp.\ 17584--17600. PMLR, 2022

    2022

  20. [28]

    Production planning by mixed integer programming, volume 149

    Yves Pochet and Laurence A Wolsey. Production planning by mixed integer programming, volume 149. Springer, 2006

    2006

  21. [29]

    Symmetry breaking using stabilizers

    Jean-Francois Puget. Symmetry breaking using stabilizers. In International Conference on Principles and Practice of Constraint Programming, pp.\ 585--599. Springer, 2003

    2003

  22. [30]

    A comparison of sbds and dynamic lex constraints

    Jean-Fran c ois Puget. A comparison of sbds and dynamic lex constraints. Symmetry and Constraint Satisfaction Problems, pp.\ 56, 2006

    2006

  23. [31]

    Solving steel mill slab problems with constraint-based techniques: Cp, lns, and cbls

    Pierre Schaus, Pascal Van Hentenryck, Jean-No \"e l Monette, Carleton Coffrin, Laurent Michel, and Yves Deville. Solving steel mill slab problems with constraint-based techniques: Cp, lns, and cbls. Constraints, 16: 0 125--147, 2011

    2011

  24. [32]

    A model for the delay management problem based on mixed-integer-programming

    Anita Sch \"o bel. A model for the delay management problem based on mixed-integer-programming. Electronic notes in theoretical computer science, 50 0 (1): 0 1--10, 2001

    2001

  25. [33]

    The bin-packing problem: A problem generator and some numerical experiments with ffd packing and mtp

    Petra Schwerin and Gerhard W \"a scher. The bin-packing problem: A problem generator and some numerical experiments with ffd packing and mtp. International transactions in operational research, 4 0 (5-6): 0 377--389, 1997

    1997

  26. [34]

    Finding symmetry breaking order parameters with euclidean neural networks

    Tess E Smidt, Mario Geiger, and Benjamin Kurt Miller. Finding symmetry breaking order parameters with euclidean neural networks. Physical Review Research, 3 0 (1): 0 L012002, 2021

    2021

  27. [35]

    Progressive hedging innovations for a class of stochastic mixed-integer resource allocation problems

    Jean-Paul Watson and David L Woodruff. Progressive hedging innovations for a class of stochastic mixed-integer resource allocation problems. Computational Management Science, 8 0 (4): 0 355--370, 2011

    2011

  28. [36]

    Equivariant symmetry breaking sets

    YuQing Xie and Tess Smidt. Equivariant symmetry breaking sets. arXiv preprint arXiv:2402.02681, 2024

    2024 arXiv

  29. [37]

    Multiset-equivariant set prediction with approximate implicit differentiation

    Yan Zhang, David W Zhang, Simon Lacoste-Julien, Gertjan J Burghouts, and Cees GM Snoek. Multiset-equivariant set prediction with approximate implicit differentiation. arXiv preprint arXiv:2111.12193, 2021

    2021 arXiv

  30. [38]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  31. [39]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  32. [40]

    B> Z B - D eg>s/J @p6 7_ _1U޿ i /_

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.