Pith. sign in

REVIEW 5 major objections 5 minor 85 references

Learning to Optimize at Scale: A Benders Decomposition-TransfORmers Framework for Stochastic Combinatorial Optimization

T0 review · 5 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A Transformer trained on 90-period lot-sizing instances, embedded in Benders decomposition with a sliding-window generator, produces feasible near-optimal solutions for 270-period stochastic instances.

desk verdict Sound hybrid with exact cuts, but the T=270 'solved' claim rests on an undefined gap reference and the paper's own future-work note. read the letter →

arxiv 2607.22550 v1 pith:WSF6OKOW submitted 2026-05-13 math.OC cs.LG

Seung Jin Choi , Kimiya Jozani , Josh Cooper , Esra Buyuktahtakin Toy This is my paper · ORCID
classification math.OCcs.LG MSC 90C1190C1568T07
keywords Bendersdecompositionstochasticcapacitatedlot-sizingtwo-stageprogrammingTransformersurrogatelearningtooptimizesliding-windowgenerationcutqualitymixed-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

The paper tries to establish that the main bottleneck of Benders decomposition—repeatedly solving scenario subproblems—can be replaced by a pre-trained Transformer that predicts binary setup decisions, leaving a small linear program to solve exactly for duals and valid cuts. It claims this hybrid reaches an average 3.55% optimality gap at the native T=90 horizon versus 41.81% for classical Benders under a 300-second limit, and stays at 2.80% even when only 50 cuts are allowed. The central scalability claim is that a sliding-window 'expandable generation' mechanism lets the same T=90-trained model handle T=270 instances zero-shot, yielding a 19.60% final gap with zero infeasibility in the generated subproblem solutions. A sympathetic reader would care because, if true, the cost of solving large two-stage stochastic programs shifts from expensive repeated exact solves to one amortized training phase plus cheap forward passes. The framework is positioned as problem-agnostic, extendable to other two-stage stochastic mixed-integer programs.

What carries the argument

The load-bearing machinery is Benders decomposition accelerated by a Transformer surrogate. The master problem selects binary setup decisions, and each scenario subproblem is normally a mixed-integer program; the Transformer instead predicts the binary setup variables in a single forward pass from a masked input tensor, reducing the subproblem to a linear program solved exactly for duals. Those duals feed the family of Benders optimality and feasibility cuts that iteratively approximate the expected recourse function. The second piece is the expandable sliding-window generator: to extend beyond the training horizon, the model generates the first window, then slides, using the tail of the pre

What would settle it

Compute a certified lower bound for the T=270 test instances (for example, the LP relaxation of the extensive form or a Benders master bound converged to tolerance) and recompute the 19.60% average gap against that bound; if the gap is substantially larger, the 'solves T=270' claim is unsupported. Also compare iteration-matched classical Benders on the same instances: if classical Benders reaches a lower final gap with the same iteration count, the Transformer surrogate is not the source of the scalability.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that a 6.8M-parameter Transformer, trained only on deterministic capacitated lot-sizing instances of horizon T=90, can act as a surrogate solver for scenario subproblems in Benders decomposition: it completes a partial master solution by predicting the binary setup variables, and once those are fixed the subproblem is an LP whose exact duals generate valid and information-dense Benders cuts. The paper reports that this ML-Benders variant closes the average optimality gap to 3.55% at T=90 (a 91.5% relative improvement over classical Benders in the same time limit), remains at 2.80% under a 50-cut budget, and extends zero-shot to T=270 via a sliding-w

Load-bearing premise

The load-bearing premise is that a Transformer trained on deterministic T=90 instances predicts setup decisions accurately enough on unseen T=270 stochastic instances that the Benders cuts built from the resulting LP duals are strong; this is supported by a reported 19.60% final gap whose reference value is not stated, while the paper notes the large-scale stochastic experiments are still being completed.

Editorial extensions

If this is right

  • If the claims hold, a single training run on small deterministic instances transfers to much larger stochastic instances, so expensive exact solving is pushed into the one-time training phase.
  • At the native horizon, the surrogate cuts close the gap to 3.55% in less time and with 29.4% fewer iterations than classical Benders, suggesting the method is useful even when scalability is not the goal.
  • The 2.80% gap under a 50-cut budget implies the learned cuts are information-dense, so the method tolerates heavily restricted cut budgets.
  • Because the framework keeps exact LP solving for duals, every generated cut is valid for the original problem, preserving the rigor of classical Benders.
  • The sliding-window mechanism implies the approach can in principle be applied to arbitrary horizons without retraining, as long as the subproblem LPs remain tractable.

Reading between the lines

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

  • The paper does not state the reference value for the 19.60% T=270 gap and says large-scale stochastic experiments are still being completed; until that gap is recomputed against a certified lower bound, a cautious reading is that the framework produces feasible, structured solutions at T=270 whose true distance from optimality is not yet certified.
  • A natural testable extension is to train the same architecture on horizons T=60 and T=120 and measure whether the zero-shot gap scales linearly, sublinearly, or abruptly degrades past a window-size threshold; that would reveal whether the sliding window or the model's internal representation is the limiting factor.
  • The training/evaluation split implies a domain-transfer claim—deterministic optimal setup patterns transfer to stochastic demand scenarios. An ablation that trains on stochastic scenarios instead would separate the value of the surrogate from the value of the domain transfer.
  • If applied to problems with integer second-stage variables, fixing the first-stage decisions will not reduce subproblems to LPs; a second-stage ML predictor or a different exact solver would be needed, so the method's current form is best matched to continuous recourse.
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 / 5 minor

Summary. The paper proposes ML-Benders, a learning-augmented Benders decomposition framework for the two-stage stochastic capacitated lot-sizing problem (TSSCLSP). A Transformer trained on T=90 deterministic lot-sizing instances is used to propose binary setup decisions; for each proposed Y, the scenario subproblem becomes an LP that is solved exactly, and the resulting duals produce valid Benders cuts for the master problem. A sliding-window generation mechanism allows the fixed-horizon model to be applied to longer horizons. The main empirical claims are: (i) on 50 T=90 instances under a 300-second limit, ML-Benders attains a 3.55% average final gap versus 41.81% for classical Benders, with fewer iterations and less time; (ii) under a strict 50-cut budget, ML-Benders achieves 2.80% versus 77.03% for classical Benders (the latter averaged over 9 instances because one failed); and (iii) the method produces feasible, complete solutions for T=270 instances, with an average final gap of 19.60% and zero subproblem infeasibility. The abstract summarizes these as solving T=270 instances at a scale previously intractable for this approach.

Significance. If the empirical claims are substantiated, the paper demonstrates a useful hybrid: the Transformer serves only as a warm-start mechanism, while exact LP duals generate the cuts, so the final solution quality is protected by the Benders convergence framework. This is a meaningful step toward scalable ML-augmented stochastic programming, and the expandable sliding-window mechanism is an interesting idea with potential transfer to other two-stage settings. The paper should be credited for using exact duals rather than letting the learned model directly determine the objective value, which avoids the circularity that plagues many learning-to-optimize papers. However, the significance cannot be fully assessed in the current form. The headline T=270 result reports a gap without a defined reference, no error bars or seeds are given for any experiment, one classical baseline instance is silently dropped, and the paper itself states that the large-scale stochastic experiments are incomplete. These are load-bearing issues for the central scalability claim.

major comments (5)
  1. [§4.1, Table 2] The headline claim that the method 'solves instances up to T=270' rests on a 'Final Opt. Gap' of 19.60%, but the reference value against which this gap is computed is never defined. In Benders decomposition, an optimality gap is typically the difference between the best incumbent (upper bound) and a master lower bound. If the master lower bound is weak, the reported gap can severely understate the true distance to optimality. Please state explicitly what the gap is measured against (exact optimal value, a documented lower bound such as an LP relaxation of the extensive form, or a Lagrangian bound), report per-instance results, the number of instances, the number of scenarios, and any solver tolerances. Without this, the T=270 'solved' claim is unsupported.
  2. [§5 and §4.1] The paper itself states in §5: 'Future work will focus on completing the large-scale stochastic experiments.' This indicates that the T=270 experiments are preliminary, yet the abstract and §4.1 present T=270 as the central demonstrated capability. Combined with the absence of instance counts, scenario counts, and any measure of variance (seeds, standard errors), the scalability claim is not yet evidenced. Please either complete the large-scale experiments or clearly label the T=270 results as preliminary and supply all experimental metadata needed to interpret them.
  3. [§4.2.2, Table 4] In Table 4, the classical Benders average of 77.03% is computed over 9 instances because 'one instance failed to find a feasible solution.' Dropping a failed instance biases the comparison in favor of ML-Benders, and the subsequent sentence 'This provides conclusive evidence' is too strong. Please report the 10-instance result including the failure (e.g., as infeasible, timeout, or a gap of 100%), or provide a sensitivity analysis showing the conclusion is unchanged under different handling of the failed instance.
  4. [§3.1–§3.2 and §4.2] The manuscript claims 'zero infeasibility in the generated subproblem solutions' and mentions that the framework generates 'optimality and feasibility cuts,' but no feasibility-cut formulation is given and Eq. (8) only displays optimality cuts. Because the Transformer proposes Y before the subproblem LP is solved, it is essential to specify what happens when a proposed Y makes a scenario subproblem infeasible: how is infeasibility detected, what dual (e.g., Farkas) certificate is used, and how is the resulting feasibility cut added to the master? Without this, the zero-infeasibility claim and the guarantee that the decomposition remains valid are not verifiable.
  5. [§3.3, §4.2] The experiments rely on several additional hyperparameters beyond the Transformer weights: 5-shot sampling with best-feasible selection, the sliding-window length, and the piecewise transform bin count. However, no random seeds, instance-generation details, or code are provided. The sampling introduces stochasticity, so the averaged gaps and iteration counts in Tables 2–4 need error bars or confidence intervals across the 50 instances (and across repeated sampling runs). Without this information, the claimed 91.5% improvement in Table 3 is not reproducible and may not be statistically significant.
minor comments (5)
  1. [Eq. (8)] The notation in the optimality cut is unclear: the dimensions of π_k and θ_k are not defined, and the cut should be written in terms of the LP dual solution and the subproblem objective value for a fixed Ŷ. Please define all symbols and clarify the index set K_ω.
  2. [§3.3, Figure 1] The 'Piecewise transforms (10 bins)' and the exact input encoding are not described. Since these affect the Transformer's ability to generalize to longer horizons, a precise description is needed for reproducibility.
  3. [Throughout] The name 'TransfORmer' is used with inconsistent capitalization; please standardize to a single spelling.
  4. [§4.2.1] The phrase 'clear superiority across all metrics' is based on a single set of averages without statistical testing or error bars. Please soften or add statistical support.
  5. [§5] The sentence 'The expandable generation mechanism ... proving that models trained on smaller ... can generalize' uses 'proving' too strongly; 'demonstrating' would be more accurate given the limited experimental evidence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the Transformer proposes Y values, but all cuts and feasibility are determined by exact LP solves; the undefined T=270 gap reference is a reporting limitation, not a circular derivation.

full rationale

The paper's derivation chain is not circular. The TransfORmer predicts binary setup decisions Y_t for scenario subproblems, but the paper states: 'With all Y_t variables fixed, the SP reduces to a fast LP, which is solved to obtain exact costs and duals. These duals are used to construct valid Benders cuts.' Thus, the final cuts and the objective bounds come from exact LP duals, not from the network's own outputs. The learning-based component only proposes candidate Y vectors; even a poor prediction would still generate valid Benders cuts, so the 'prediction' is not the fitted answer but a warm start. The expandable sliding-window mechanism is an inference-time procedure that does not feed T=270 data into training; the model is explicitly trained only on T=90 deterministic instances. The comparison against classical Benders under the same cut budget is an external, falsifiable benchmark rather than a self-referential validation. The only notable issue is that Table 2 reports a 19.60% 'Final Opt. Gap' for T=270 without defining the reference value, and Section 5 states 'Future work will focus on completing the large-scale stochastic experiments.' This is a serious reporting/verification gap for the headline scalability claim, but it is not a case where a result is equivalent to its inputs by construction. Self-citations to the authors' prior expandable ML-optimization work appear in the introduction and references, but they are contextual background, not the load-bearing justification for the paper's central results. Consequently, no circular step meeting the specified evidentiary standard is present.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on four fitted design choices (model weights, sampling/selection, window length, bin count) and four assumptions, two of which are ad hoc or unverified. No new physical entities are introduced. The largest burden is the unstated T=270 gap reference, which makes the headline scalability claim unverifiable from the provided text.

free parameters (4)
  • Transformer weights = 6.8M parameters
    The entire cut-quality claim rests on these learned weights; trained on deterministic T=90 CLSP instances. Location: Table 1.
  • 5-shot sampling with best-feasible selection = 5 samples; best feasible chosen
    Chosen procedure that guarantees 'zero infeasibility' by construction; if no feasible sample exists the behavior is unstated. Location: §3.3.
  • Sliding window length = Presumably 90 (training horizon)
    The window size controls how the model generalizes to T=270; not varied or justified. Location: §3.3.
  • Piecewise transform bin count = 10
    Hand-chosen quantization for input features; affects prediction quality. Location: Figure 1 caption.
assumptions (4)
  • standard math Benders decomposition correctness: dual solutions of scenario LPs generate valid optimality and feasibility cuts.
    Invoked in §3.1 without proof; standard result, not a weakness.
  • domain assumption The deterministic T=90 CLSP training distribution is representative of the stochastic scenario subproblems encountered at T=90 and T=270.
    The model is trained 'exclusively on deterministic ... T=90' and used on stochastic instances; no distributional analysis is provided. Location: §4.
  • domain assumption The T=270 'optimality gap' is computed against a valid reference value (exact optimal or a strong bound).
    No reference for the gap is given; without it the 19.60% number is uninterpretable. Location: Table 2.
  • ad hoc to paper Sliding-window concatenation of the previous output as historical context preserves solution validity for horizons far beyond training.
    The sliding-window heuristic is asserted to work for T=270 but no correctness argument or failure analysis is given. Location: §3.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning to Optimize at Scale: A Benders Decomposition-TransfORmers Framework for Stochastic Combinatorial Optimization." pith.science (2026). https://pith.science/paper/WSF6OKOW

@misc{pith2026260722550,
  author       = {Pith},
  title        = {Pith review of: Learning to Optimize at Scale: A Benders Decomposition-TransfORmers Framework for Stochastic Combinatorial Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WSF6OKOW}},
  note         = {Machine review of arXiv:2607.22550}
}
read the original abstract

We propose a learning-augmented Benders decomposition framework to solve large-scale two-stage stochastic mixed-integer programs. We focus on the two-stage stochastic capacitated lot-sizing problem (TSSCLSP) under demand uncertainty. Our method accelerates the convergence of the decomposition by using a pre-trained TransfORmer model to rapidly generate high-quality approximate solutions for the scenario subproblems. This hybrid strategy uses the TransfORmer predictions to generate strong optimality and feasibility cuts, effectively guiding the Benders master problem. Our framework includes a novel expandable generation mechanism, allowing a model trained on a fixed horizon to solve instances of arbitrary length. For the test set considered, our method solves instances up to T = 270, a scale previously intractable for this approach, while maintaining zero infeasibility in the generated subproblem solutions. This demonstrates the potential of TransfORmers as powerful surrogate solvers embedded within classical decomposition algorithms.

Figures

Figures reproduced from arXiv: 2607.22550 by the authors.

Figure 1
Figure 1. TransfORmer architecture used to solve subproblems in the decomposition framework. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

85 extracted references · 1 canonical work pages

  1. [1]

    and Mingozzi, A

    Baldacci, R. and Mingozzi, A. and Roberti, R. , title =. Naval Research Logistics , volume =. 2009 , doi =

    2009

  2. [2]

    and Poggi, M

    Pecin, D. and Poggi, M. and Uchoa, E. , title =. INFORMS Journal on Computing , volume =. 2022 , doi =

    2022

  3. [3]

    and Kirkpatrick, B

    Diaz, N. and Kirkpatrick, B. and Pistikopoulos, E.N. , title =. IFAC-PapersOnLine , volume =. 2022 , doi =

    2022

  4. [4]

    and Tsakiris, I

    Bakos, Y. and Tsakiris, I. and Konstantas, G. , title =. International Journal of Production Economics , volume =. 1998 , doi =

    1998

  5. [5]

    and Wang, J

    Zhang, W. and Wang, J. and Li, Y. , title =. arXiv preprint , volume =. 2024 , url =

    2024

  6. [6]

    and Kosmatos, A

    Faísca, N.P. and Kosmatos, A. and Dua, V. and Pistikopoulos, E.N. , title =. Journal of Optimization Theory and Applications , volume =. 2014 , doi =

    2014

  7. [7]

    Management Science , volume=

    A dynamic lot-sizing model with demand time windows , author=. Management Science , volume=. 2001 , publisher=

    2001

  8. [8]

    Annals of Operations Research , volume=

    Stage-t scenario dominance for risk-averse multi-stage stochastic mixed-integer programs , author=. Annals of Operations Research , volume=. 2022 , publisher=

    2022

Show all 85 references
  1. [9]

    European Journal of Operational Research , volume=

    The Benders decomposition algorithm: A literature review , author=. European Journal of Operational Research , volume=. 2017 , publisher=

    2017

  2. [10]

    Operations Research Forum , volume=

    Learning optimal solutions via an LSTM-optimization framework , author=. Operations Research Forum , volume=. 2023 , organization=

    2023

  3. [11]

    Computers & Operations Research , volume=

    Scenario-dominance to multi-stage stochastic lot-sizing and knapsack problems , author=. Computers & Operations Research , volume=. 2023 , publisher=

    2023

  4. [12]

    nature , volume=

    Deep learning , author=. nature , volume=. 2015 , publisher=

    2015

  5. [13]

    2016 , publisher=

    Deep learning , author=. 2016 , publisher=

    2016

  6. [14]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Transformer-based learned optimization , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  7. [15]

    Advances in Neural Information Processing Systems , volume=

    Transformers from an optimization perspective , author=. Advances in Neural Information Processing Systems , volume=

  8. [16]

    IISE Annual Conference

    Toward TransfORmers: Revolutionizing the Solution of Mixed Integer Programs with Transformers , author=. IISE Annual Conference. Proceedings , pages=. 2024 , organization=

    2024

  9. [17]

    Mathematical Methods in the Social Sciences , year=

    The optimality of (s, S) policies in the dynamic inventory problem , author=. Mathematical Methods in the Social Sciences , year=

  10. [18]

    2000 , publisher=

    Foundations of inventory management , author=. 2000 , publisher=

    2000

  11. [19]

    Management Science , volume=

    Strategies for the probabilistic lot-sizing problem with service-level constraints , author=. Management Science , volume=. 1988 , publisher=

    1988

  12. [20]

    Advances in neural information processing systems , pages=

    Attention is all you need , author=. Advances in neural information processing systems , pages=

  13. [21]

    Computers & Industrial Engineering , volume=

    Enhancing the accuracy and interpretability of neural networks for demand forecasting in supply chain management , author=. Computers & Industrial Engineering , volume=. 2019 , publisher=

    2019

  14. [22]

    Philosophical Transactions of the Royal Society A , volume=

    Time-series forecasting with deep learning: a survey , author=. Philosophical Transactions of the Royal Society A , volume=. 2021 , publisher=

    2021

  15. [23]

    Foundations and Trends in Machine Learning , year=

    Transformers for time series forecasting in supply chain systems: A survey , author=. Foundations and Trends in Machine Learning , year=

  16. [24]

    2009 , publisher=

    Lectures on Stochastic Programming: Modeling and Theory , author=. 2009 , publisher=

    2009

  17. [25]

    Operations Research , volume=

    Dynamic lot sizing with random demand and nonstationary costs , author=. Operations Research , volume=. 1997 , publisher=

    1997

  18. [26]

    AIIE Transactions , volume=

    A procedure for production lot sizing with probabilistic dynamic demand , author=. AIIE Transactions , volume=

  19. [27]

    Operations Research , volume=

    Decomposition and partitioning methods for multistage stochastic linear programs , author=. Operations Research , volume=

  20. [28]

    Computers & Chemical Engineering , volume=

    A combined MILP and CP decomposition method for multi-period production planning problems , author=. Computers & Chemical Engineering , volume=

  21. [29]

    European Journal of Operational Research , volume=

    Benders decomposition algorithm: A literature review , author=. European Journal of Operational Research , volume=

  22. [30]

    European Journal of Operational Research , volume=

    Machine learning for combinatorial optimization: a methodological tour d'horizon , author=. European Journal of Operational Research , volume=

  23. [31]

    Advances in Neural Information Processing Systems , volume=

    Learning combinatorial optimization algorithms over graphs , author=. Advances in Neural Information Processing Systems , volume=

  24. [32]

    Computers & Industrial Engineering , volume=

    A deep reinforcement learning approach for stochastic lot-sizing and scheduling problems , author=. Computers & Industrial Engineering , volume=

  25. [33]

    2024 , issn =

    Deep learning model for optimizing control and planning in stochastic manufacturing environments , journal =. 2024 , issn =. doi:https://doi.org/10.1016/j.eswa.2024.125075 , author =

    2024

  26. [34]

    1999 , issn =

    A review of the stochastic lot scheduling problem1This paper is based upon work supported in part by the National Science Foundation under Grant Number DMI-9409344.1 , journal =. 1999 , issn =. doi:https://doi.org/10.1016/S0925-5273(98)00247-3 , author =

    1999 doi

  27. [35]

    European Journal of Operational Research , volume=

    Single item lot sizing problems , author=. European Journal of Operational Research , volume=. 2006 , publisher=

    2006

  28. [36]

    J. F. Benders , title =. Numerische Mathematik , year =

  29. [37]

    R. M. Van Slyke and R. J.-B. Wets , title =. SIAM Journal on Applied Mathematics , year =

  30. [38]

    R. T. Rockafellar and R. J.-B. Wets , title =. Mathematics of Operations Research , year =

  31. [39]

    G. B. Dantzig and P. Wolfe , title =. Operations Research , year =

  32. [40]

    Buschk\"uhl and F

    L. Buschk\"uhl and F. Sahling and S. Helber and H. Tempelmeier , title =. OR Spectrum , year =

  33. [41]

    Gruson and J.-F

    M. Gruson and J.-F. Cordeau and R. Jans , title =. European Journal of Operational Research , year =

  34. [42]

    K. K. Haugen and A. L kketangen and D. L. Woodruff , title =. European Journal of Operational Research , year =

  35. [43]

    J. R. Birge and F. V. Louveaux , title =. 2011 , edition =

    2011

  36. [44]

    Heitsch and W

    H. Heitsch and W. R\"omisch , title =. Computational Optimization and Applications , year =

  37. [45]

    A. J. Kleywegt and A. Shapiro and T. Homem-de-Mello , title =. SIAM Journal on Optimization , year =

  38. [46]

    Gong and Y

    H. Gong and Y. Zhang and Z.-H. Zhang , title =. IEEE Transactions on Automation Science and Engineering (submitted) , year =

  39. [47]

    J. K. A Machine Learning Approach for Predicting the Best Heuristic for a Large Scaled Capacitated Lot Sizing Problem , journal =. 2025 , note =

    2025

  40. [48]

    Schlenkrich and J.-F

    M. Schlenkrich and J.-F. Cordeau and S. N. Parragh , title =. arXiv preprint arXiv:2503.08477 , year =

    arXiv

  41. [49]

    2020 , publisher=

    Integer programming , author=. 2020 , publisher=

    2020

  42. [50]

    Handbooks in operations research and management science , volume=

    Algorithms for stochastic mixed-integer programming models , author=. Handbooks in operations research and management science , volume=. 2005 , publisher=

    2005

  43. [51]

    Woodruff , title =

    Jean-Paul Watson and David L. Woodruff , title =. Computational Management Science , volume =. 2011 , publisher =

    2011

  44. [52]

    Van Slyke, R. M. and Wets, R. J.-B. , title =. SIAM Journal on Applied Mathematics , year =

  45. [53]

    Birge, J. R. and Louveaux, F. , title =

  46. [54]

    Magnanti, T. L. and Wong, R. T. , title =. Operations Research , year =

  47. [55]

    Geoffrion, A. M. , title =. Journal of Optimization Theory and Applications , year =

  48. [56]

    and Fischetti, M

    Codato, G. and Fischetti, M. , title =. Operations Research , year =

  49. [57]

    and Cordeau, J.-F

    Contreras, I. and Cordeau, J.-F. and Laporte, G. , title =. Operations Research , year =

  50. [58]

    Rockafellar, R. T. and Wets, R. J.-B. , title =. Mathematics of Operations Research , year =

  51. [59]

    Pereira, M. V. F. and Pinto, L. M. V. G. , title =. Journal of Optimization Theory and Applications , year =

  52. [60]

    and Dash, S

    Bodur, M. and Dash, S. and Günlük, O. and Luedtke, J. , title =. INFORMS Journal on Computing , year =

  53. [61]

    Dantzig, G. B. and Wolfe, P. , title =. Operations Research , year =

  54. [62]

    and Saharidis, G

    Fragkogiannis, A. and Saharidis, G. K. , title =. Encyclopedia of Information Science and Technology, 4th ed. , year =

  55. [63]

    and Ahmed, S

    Rahmaniani, R. and Ahmed, S. and Crainic, T. G. and Gendreau, M. and Rei, W. , title =. arXiv preprint arXiv:1702.05633 , year =

    arXiv

  56. [64]

    Mathematical programming , volume=

    On solving discrete two-stage stochastic programs having mixed-integer first-and second-stage variables , author=. Mathematical programming , volume=. 2006 , publisher=

    2006

  57. [65]

    International Journal of Production Research , volume=

    Using the proximal policy optimisation algorithm for solving the stochastic capacitated lot sizing problem , author=. International Journal of Production Research , volume=. 2023 , doi=

    2023

  58. [66]

    Expert Systems with Applications , volume=

    Reinforcement learning approaches for the stochastic discrete lot-sizing problem on parallel machines , author=. Expert Systems with Applications , volume=. 2024 , pages=

    2024

  59. [67]

    European Journal of Operational Research , volume=

    Deep Controlled Learning for Inventory Control , author=. European Journal of Operational Research , volume=. 2025 , doi=

    2025

  60. [68]

    INFORMS Journal on Optimization , volume=

    Benders cut classification via support vector machines for solving two-stage stochastic programs , author=. INFORMS Journal on Optimization , volume=. 2021 , doi=

    2021

  61. [69]

    IEEE Transactions on Wireless Communications , year=

    Accelerating generalized Benders decomposition for wireless resource allocation , author=. IEEE Transactions on Wireless Communications , year=

  62. [70]

    Proceedings of FOCAPO 2023 / CPC Conference , year=

    Learning to initialize generalized Benders decomposition via active learning , author=. Proceedings of FOCAPO 2023 / CPC Conference , year=

    2023

  63. [71]

    Journal of Process Control , volume=

    Computationally efficient solution of mixed integer model predictive control problems via machine learning aided Benders decomposition , author=. Journal of Process Control , volume=. 2024 , doi=

    2024

  64. [72]

    Proceedings of the 38th Conference on Uncertainty in Artificial Intelligence , year=

    Neural-Progressive Hedging: Enforcing Constraints in Reinforcement Learning with Stochastic Programming , author=. Proceedings of the 38th Conference on Uncertainty in Artificial Intelligence , year=

  65. [73]

    arXiv preprint arXiv:2302.09166 , year=

    Machine learning for cutting planes in integer programming: A survey , author=. arXiv preprint arXiv:2302.09166 , year=

    arXiv

  66. [74]

    Operations Research Letters , volume=

    The integer L-shaped method for stochastic integer programs with complete recourse , author=. Operations Research Letters , volume=. 1993 , publisher=

    1993

  67. [75]

    Operations Research , volume=

    Combinatorial Benders' cuts for mixed-integer linear programming , author=. Operations Research , volume=. 2006 , publisher=

    2006

  68. [76]

    Constraint Programming and Large Scale Discrete Optimization , pages=

    A hybrid method for planning and scheduling , author=. Constraint Programming and Large Scale Discrete Optimization , pages=. 2005 , publisher=

    2005

  69. [77]

    Naval Research Logistics (NRL) , volume=

    An integer decomposition algorithm for solving a two-stage facility location problem with second-stage activation costs , author=. Naval Research Logistics (NRL) , volume=. 2010 , publisher=

    2010

  70. [78]

    Mathematical Programming , volume=

    Logic-based Benders decomposition , author=. Mathematical Programming , volume=. 2003 , publisher=

    2003

  71. [79]

    Operations Research , volume=

    An integer L-shaped algorithm for the capacitated vehicle routing problem with stochastic demands , author=. Operations Research , volume=. 2002 , publisher=

    2002

  72. [80]

    Mathematical Programming , volume=

    A study of the lot-sizing polytope , author=. Mathematical Programming , volume=. 2004 , publisher=

    2004

  73. [81]

    European Journal of Operational Research , volume=

    An expandable machine learning-optimization framework to sequential decision-making , author=. European Journal of Operational Research , volume=. 2024 , publisher=

    2024

  74. [82]

    Optimization Letters , volume=

    A deep reinforcement learning framework for solving two-stage stochastic programs , author=. Optimization Letters , volume=. 2024 , publisher=

    2024

  75. [83]

    Annals of Operations Research , pages=

    A non-anticipative learning-optimization framework for solving multi-stage stochastic programs , author=. Annals of Operations Research , pages=. 2024 , publisher=

    2024

  76. [84]

    Journal of Global Optimization , volume=

    A K-means supported reinforcement learning framework to multi-dimensional knapsack , author=. Journal of Global Optimization , volume=. 2024 , publisher=

    2024

  77. [85]

    Annals of Operations Research , volume=

    A simulation-deep reinforcement learning (SiRL) approach for epidemic control optimization , author=. Annals of Operations Research , volume=. 2023 , publisher=

    2023

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.