pith. machine review for the scientific record. sign in

arxiv: 2604.22107 · v1 · submitted 2026-04-23 · 📡 eess.SY · cs.SY

Recognition: unknown

A Hybrid Reinforcement and Self-Supervised Learning Aided Benders Decomposition Algorithm

Bernard T. Agyeman, Ilias Mitrai, Prodromos Daoutidis, Zhe Li

Authors on Pith no claims yet

Pith reviewed 2026-05-09 20:26 UTC · model grok-4.3

classification 📡 eess.SY cs.SY
keywords generalized Benders decompositionreinforcement learningself-supervised learningmixed integer nonlinear programmingKKT conditionsneural networksoptimization
0
0 comments X

The pith

A graph-based reinforcement learning agent and KKT-informed neural network accelerate generalized Benders decomposition.

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

The authors develop a hybrid learning method to make generalized Benders decomposition faster for solving mixed-integer nonlinear optimization problems. Instead of solving the master problem and subproblem repeatedly with traditional solvers, a reinforcement learning agent learns to choose good integer assignments for the master problem using a graph representation and a check for validity. These choices then go to a neural network that has learned, without labeled data, to guess the primal and dual solutions of the subproblem that nearly satisfy its optimality conditions. The guesses allow the algorithm to build the necessary cutting planes right away. On a test case the new method cuts the total time to solution by 57.5 percent while still reaching the best possible answer every time.

Core claim

A graph based reinforcement learning agent operates on a bipartite representation of the master problem and, together with a verification mechanism, determines the integer variable assignments that solve the master problem. These assignments are then used as inputs to a KKT informed neural network, trained via self supervision to predict primal dual solutions that approximately satisfy the Karush Kuhn Tucker conditions of the subproblem. The predicted solutions are used to construct Benders cuts directly. When evaluated on a mixed integer nonlinear programming case study, the framework achieves a 57.5% reduction in solution time relative to classical GBD while consistently recovering optimal

What carries the argument

Graph-based reinforcement learning agent on bipartite master problem representation with verification, paired with a self-supervised KKT-informed neural network that predicts approximate primal-dual subproblem solutions for direct Benders cut generation.

If this is right

  • The master problem is solved by learning rather than repeated optimization calls.
  • Valid Benders cuts can be generated from approximate rather than exact subproblem solutions.
  • The overall algorithm converges to the same optimal solutions as classical GBD.
  • Solution time is reduced by more than half on the evaluated MINLP instances.

Where Pith is reading between the lines

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

  • This approach may extend to other cutting-plane or decomposition methods in optimization.
  • Self-supervised training could enable the method to adapt online to varying problem parameters.
  • Success depends on the verification step preventing bad integer assignments from the RL agent.
  • Further speedups might come from integrating the two learning components more tightly.

Load-bearing premise

Approximate satisfaction of the KKT conditions by the neural network predictions is sufficient to produce Benders cuts that are valid and lead to convergence to the optimum.

What would settle it

If the hybrid algorithm produces final solutions worse than the known optimum or if the cuts it generates are invalid on the test instances, causing failure to converge or longer run times.

Figures

Figures reproduced from arXiv: 2604.22107 by Bernard T. Agyeman, Ilias Mitrai, Prodromos Daoutidis, Zhe Li.

Figure 1
Figure 1. Figure 1: Workflow of the graph-based agent for the master problem. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The architecture of the proposed KINN. Algorithm 1 KINN training procedure Require: Y = {y i} N i=1, KINNω, ℓ, epochs T, batch size B, learning rate η 1: for t = 1, 2, . . . , T do 2: for each mini-batch YB ⊂ Y of size B do 3: For each y i ∈ YB, compute (xˆ i , µˆi ,λˆ i) ← KINNω(y i ) 4: Evaluate L(ω) ← 1 |YB| P yi∈YB ℓ(y i , xˆ i , µˆi ,λˆ i) 5: g ← ∇ωL(ω) 6: ω ← ω − η g 7: end for 8: end for 9: return ω… view at source ↗
read the original abstract

We propose a hybrid reinforcement and self-supervised learning framework for accelerating generalized Benders decomposition (GBD). In this framework, a graph based reinforcement learning agent operates on a bipartite representation of the master problem and, together with a verification mechanism, determines the integer variable assignments that solve the master problem. These assignments are then used as inputs to a KKT informed neural network, trained via self supervision to predict primal dual solutions that approximately satisfy the Karush Kuhn Tucker conditions of the subproblem. The predicted solutions are used to construct Benders cuts directly. The framework is evaluated on a mixed integer nonlinear programming case study, where it achieves a 57.5% reduction in solution time relative to classical GBD while consistently recovering optimal solutions across all test instances.

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

3 major / 2 minor

Summary. The manuscript proposes a hybrid reinforcement and self-supervised learning framework to accelerate generalized Benders decomposition (GBD) for mixed-integer nonlinear programs. A graph-based RL agent, operating on a bipartite representation of the master problem together with a verification mechanism, generates integer variable assignments. These assignments are passed to a KKT-informed neural network trained via self-supervision to predict approximate primal-dual solutions for the subproblem; the predictions are inserted directly into the standard Benders cut formula. On an MINLP case study the method reports a 57.5% reduction in solution time relative to classical GBD while recovering optimal solutions on all test instances.

Significance. If the generated cuts remain valid and the approach generalizes beyond the single case study, the hybrid RL-NN acceleration could meaningfully reduce the computational burden of decomposition methods for large-scale MINLPs by limiting the number of exact subproblem solves. The combination of graph RL for discrete master decisions and self-supervised KKT-informed prediction for continuous subproblems is technically interesting and the reported empirical speed-up is substantial; however, the absence of cut-validity guarantees or residual bounds limits the immediate theoretical impact.

major comments (3)
  1. [abstract and method framework] The central construction (abstract and method description) inserts primal-dual pairs that only approximately satisfy the KKT conditions of the subproblem into the standard Benders optimality-cut formula. Because the dual multipliers produced by the network need not be feasible for the subproblem dual at the fixed integer point, the resulting cut is not guaranteed to be a valid supporting hyperplane of the recourse function. The manuscript provides no residual bounds, feasibility certificates, or proof that the attained approximation level preserves cut validity; the reported consistent recovery of optima therefore rests on an unverified empirical property.
  2. [abstract and RL component description] The verification mechanism that is claimed to ensure the graph-based RL agent solves the master problem exactly is described only at a high level (abstract). No details are given on its implementation, computational cost, or formal guarantee that the integer assignments passed to the NN are globally optimal for the current master problem; without this, the overall algorithm's correctness cannot be assessed.
  3. [training procedure and numerical results] The self-supervised training procedure for the KKT-informed network (loss penalizing stationarity, primal/dual feasibility, and complementarity residuals) is not accompanied by any analysis of how the attained residual level affects cut strength or validity across the range of integer assignments encountered during decomposition. Table or figure reporting residual statistics versus cut validity or iteration count would be required to substantiate the 57.5% speedup claim.
minor comments (2)
  1. [method section] Notation for the bipartite graph representation of the master problem and for the KKT residual loss should be introduced with explicit definitions and consistent symbols throughout.
  2. [numerical experiments] The MINLP case study should report instance sizes, number of Benders iterations, and a direct comparison of cut quality (e.g., lower-bound tightness) between the learned cuts and classical cuts.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which help clarify the theoretical and empirical aspects of our hybrid RL-self-supervised framework for accelerating GBD. We address each major comment point by point below and indicate the revisions we will make to the manuscript.

read point-by-point responses

  1. Referee: The central construction (abstract and method description) inserts primal-dual pairs that only approximately satisfy the KKT conditions of the subproblem into the standard Benders optimality-cut formula. Because the dual multipliers produced by the network need not be feasible for the subproblem dual at the fixed integer point, the resulting cut is not guaranteed to be a valid supporting hyperplane of the recourse function. The manuscript provides no residual bounds, feasibility certificates, or proof that the attained approximation level preserves cut validity; the reported consistent recovery of optima therefore rests on an unverified empirical property.

    Authors: We acknowledge that the KKT-informed network produces approximate primal-dual solutions and that the manuscript does not supply formal residual bounds or a proof of cut validity. The reported recovery of optimal solutions across all instances is indeed an empirical observation. In the revised manuscript we will add a new subsection in the method and results sections that (i) explicitly states the lack of theoretical guarantees, (ii) reports the observed residual statistics (stationarity, feasibility, and complementarity) for every generated cut, and (iii) discusses the practical conditions under which the approximations have preserved convergence in the tested MINLP. This will make the empirical nature of the correctness claim transparent. revision: partial

  2. Referee: The verification mechanism that is claimed to ensure the graph-based RL agent solves the master problem exactly is described only at a high level (abstract). No details are given on its implementation, computational cost, or formal guarantee that the integer assignments passed to the NN are globally optimal for the current master problem; without this, the overall algorithm's correctness cannot be assessed.

    Authors: The verification step invokes a standard MIP solver on the current master problem after the RL agent proposes an integer assignment; only assignments that are certified optimal by the solver are forwarded to the neural network. This provides a formal guarantee that the master problem is solved exactly at every iteration. We will expand the algorithm description (Section 3) with pseudocode, the specific solver and tolerance settings used, and a short complexity discussion showing that the verification overhead is negligible compared with the subproblem solves that are avoided. These additions will allow readers to assess the overall correctness. revision: yes

  3. Referee: The self-supervised training procedure for the KKT-informed network (loss penalizing stationarity, primal/dual feasibility, and complementarity residuals) is not accompanied by any analysis of how the attained residual level affects cut strength or validity across the range of integer assignments encountered during decomposition. Table or figure reporting residual statistics versus cut validity or iteration count would be required to substantiate the 57.5% speedup claim.

    Authors: We agree that a quantitative link between residual levels and cut quality would strengthen the empirical claims. In the revised version we will add a new table (and accompanying figure) that reports, for each test instance and across all decomposition iterations, the mean and maximum residuals achieved by the network together with the number of iterations required for convergence and the final objective gap. A short discussion will relate these residuals to the observed 57.5 % speedup, thereby providing the requested substantiation while remaining within the scope of the current case study. revision: yes

Circularity Check

0 steps flagged

No circularity: hybrid learning approximates external GBD steps without self-referential reduction

full rationale

The framework trains a graph RL agent (with verification) to solve the master problem and a self-supervised NN to output approximate primal-dual pairs for the subproblem by penalizing KKT residuals. These outputs are then inserted into the standard Benders cut formula. Neither component is defined in terms of the other or of the final cuts; training targets (KKT stationarity/feasibility/complementarity) are independent of the cut validity claim, which is supported only by empirical recovery of optima on the MINLP instances. No equation reduces to its input by construction, no fitted parameter is relabeled as a prediction of a related quantity, and no load-bearing premise rests on self-citation. The derivation therefore remains self-contained against external optimization benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

Based solely on the abstract, the central claim rests on standard GBD structure plus learned approximations whose parameters are fitted during training; no new physical entities are introduced.

free parameters (2)
  • Neural network parameters
    Weights and biases of the KKT-informed neural network are fitted during self-supervised training to approximate subproblem solutions.
  • Reinforcement learning policy parameters
    Parameters of the graph-based RL agent are learned to select integer assignments for the master problem.
axioms (2)
  • domain assumption A bipartite graph representation of the master problem is suitable for RL processing.
    Invoked to enable the RL agent to operate on the master problem structure.
  • ad hoc to paper Approximate satisfaction of KKT conditions by predicted solutions is sufficient to generate valid Benders cuts.
    Central to using NN predictions directly without exact subproblem solves.

pith-pipeline@v0.9.0 · 5434 in / 1615 out tokens · 37953 ms · 2026-05-09T20:26:15.972151+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

113 extracted references · 8 canonical work pages · 2 internal anchors

  1. [1]

    Partitioning procedures for solving mixed-variables programming problems , author=. Numer. Math , volume=

  2. [2]

    Electric Power Systems Research , volume=

    Physics-informed neural networks for ac optimal power flow , author=. Electric Power Systems Research , volume=. 2022 , publisher=

    2022

  3. [3]

    Learning continuous

    Warrington, Joseph , booktitle=. Learning continuous. 2019 , organization=

    2019

  4. [4]

    Inexact cuts in

    Zakeri, Golbon and Philpott, Andrew B and Ryan, David M , journal=. Inexact cuts in. 2000 , publisher=

    2000

  5. [5]

    arXiv preprint arXiv:2107.10201 , year=

    Learning a large neighborhood search algorithm for mixed integer programs , author=. arXiv preprint arXiv:2107.10201 , year=

    work page arXiv

  6. [6]

    Computers & Chemical Engineering , volume=

    Deep reinforcement learning control of hydraulic fracturing , author=. Computers & Chemical Engineering , volume=. 2021 , publisher=

    2021

  7. [7]

    Reinforcement learning based optimal control of batch processes using

    Yoo, Haeun and Kim, Boeun and Kim, Jong Woo and Lee, Jay H , journal=. Reinforcement learning based optimal control of batch processes using. 2021 , publisher=

    2021

  8. [8]

    Learning to cut by looking ahead:

    Paulus, Max B and Zarpellon, Giulia and Krause, Andreas and Charlin, Laurent and Maddison, Chris , booktitle=. Learning to cut by looking ahead:. 2022 , organization=

    2022

  9. [9]

    arXiv preprint arXiv:2206.14987 , year=

    Lookback for learning to branch , author=. arXiv preprint arXiv:2206.14987 , year=

    work page arXiv

  10. [10]

    Algorithm runtime prediction:

    Hutter, Frank and Xu, Lin and Hoos, Holger H and Leyton-Brown, Kevin , journal=. Algorithm runtime prediction:. 2014 , publisher=

    2014

  11. [11]

    Computers & Chemical Engineering , volume=

    Improvements for decomposition based methods utilized in the development of multi-scale energy systems , author=. Computers & Chemical Engineering , volume=. 2023 , publisher=

    2023

  12. [12]

    Deterministic electric power infrastructure planning:

    Lara, Cristiana L and Mallapragada, Dharik S and Papageorgiou, Dimitri J and Venkatesh, Aranya and Grossmann, Ignacio E , journal=. Deterministic electric power infrastructure planning:. 2018 , publisher=

    2018

  13. [13]

    Design and operation of modular biorefinery supply chain under uncertainty using generalized

    Luo, Yuqing and Ierapetritou, Marianthi , journal=. Design and operation of modular biorefinery supply chain under uncertainty using generalized. 2024 , publisher=

    2024

  14. [14]

    Decomposition techniques in mathematical programming:

    Conejo, Antonio J and Castillo, Enrique and Minguez, Roberto and Garcia-Bertrand, Raquel , year=. Decomposition techniques in mathematical programming:

  15. [15]

    Fast continuous and integer

    Larsen, Eric and Frejinger, Emma and Gendron, Bernard and Lodi, Andrea , journal=. Fast continuous and integer. 2023 , publisher=

    2023

  16. [16]

    Computationally efficient solution of mixed integer model predictive control problems via machine learning aided

    Mitrai, Ilias and Daoutidis, Prodromos , journal=. Computationally efficient solution of mixed integer model predictive control problems via machine learning aided. 2024 , publisher=

    2024

  17. [17]

    AIChE Journal , volume=

    A decoupling strategy for protecting sensitive process information in cooperative optimization of power flow , author=. AIChE Journal , volume=. 2022 , publisher=

    2022

  18. [18]

    Improving

    Saharidis, Georgios KD and Ierapetritou, Marianthi G , journal=. Improving. 2010 , publisher=

    2010

  19. [19]

    Computational strategies for improved

    Su, Lijie and Tang, Lixin and Grossmann, Ignacio E , journal=. Computational strategies for improved. 2015 , publisher=

    2015

  20. [20]

    Initialization of the

    Saharidis, Georgios KD and Boile, Maria and Theofanis, Sotiris , journal=. Initialization of the. 2011 , publisher=

    2011

  21. [21]

    Accelerating

    Magnanti, Thomas L and Wong, Richard T , journal=. Accelerating. 1981 , publisher=

    1981

  22. [22]

    Integrating operations and control:

    Daoutidis, Prodromos and Lee, Jay H and Harjunkoski, Iiro and Skogestad, Sigurd and Baldea, Michael and Georgakis, Christos , journal=. Integrating operations and control:. 2018 , publisher=

    2018

  23. [23]

    Computers & Chemical Engineering , volume=

    Process systems engineering--the generation next? , author=. Computers & Chemical Engineering , volume=. 2021 , publisher=

    2021

  24. [24]

    Geometric deep learning:

    Bronstein, Michael M and Bruna, Joan and LeCun, Yann and Szlam, Arthur and Vandergheynst, Pierre , journal=. Geometric deep learning:. 2017 , publisher=

    2017

  25. [25]

    Computers & Chemical Engineering , volume=

    Graph-based modeling and simulation of complex systems , author=. Computers & Chemical Engineering , volume=. 2019 , publisher=

    2019

  26. [26]

    INFORMS Journal on Computing , volume=

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

    2021

  27. [27]

    Advances in

    Learning combinatorial optimization algorithms over graphs , author=. Advances in

  28. [28]

    Proceedings of the

    Accelerating primal solution findings for mixed integer programs based on solution prediction , author=. Proceedings of the

  29. [29]

    Transportation Science , volume=

    Machine-learning--based column selection for column generation , author=. Transportation Science , volume=. 2021 , publisher=

    2021

  30. [30]

    A multicut generalized

    Mitrai, Ilias and Daoutidis, Prodromos , journal=. A multicut generalized. 2022 , publisher=

    2022

  31. [31]

    Numerische Mathematik , volume=

    Partitioning procedures for solving mixed-variables programming problems , author=. Numerische Mathematik , volume=. 1962 , publisher=

    1962

  32. [32]

    Aardal, Karen and Larsson, Torbj. A. European Journal of Operational Research , volume=. 1990 , publisher=

    1990

  33. [33]

    , author=

    Partitioning procedures for solving mixed-variables programming problems. , author=. Computational Management Science , volume=

  34. [34]

    Machine learning for combinatorial optimization:

    Bengio, Yoshua and Lodi, Andrea and Prouvost, Antoine , journal=. Machine learning for combinatorial optimization:. 2021 , publisher=

    2021

  35. [35]

    Accelerating process control and optimization via machine learning:

    Mitrai, Ilias and Daoutidis, Prodromos , journal=. Accelerating process control and optimization via machine learning:. 2025 , publisher=

    2025

  36. [36]

    Machine learning-enhanced

    Borozan, Stefan and Giannelos, Spyros and Falugi, Paola and Moreira, Alexandre and Strbac, Goran , journal=. Machine learning-enhanced. 2024 , publisher=

    2024

  37. [37]

    Accelerating

    Hasan, Fouad and Kargarian, Amin , journal=. Accelerating. 2024 , publisher=

    2024

  38. [38]

    Generalized

    Geoffrion, Arthur M , journal=. Generalized. 1972 , publisher=

    1972

  39. [39]

    , author=

    Learning Fast Optimizers for Contextual Stochastic Integer Programs. , author=. UAI , pages=

  40. [40]

    Neural Combinatorial Optimization with Reinforcement Learning

    Neural combinatorial optimization with reinforcement learning , author=. arXiv preprint arXiv:1611.09940 , year=

    work page Pith review arXiv

  41. [41]

    Accelerate hybrid model predictive control using generalized

    Lin, Xuan , journal=. Accelerate hybrid model predictive control using generalized

  42. [42]

    Rahmaniani, Ragheb and Crainic, Teodor Gabriel and Gendreau, Michel and Rei, Walter , journal=. The. 2017 , publisher=

    2017

  43. [43]

    Advances in Neural Information Processing Systems , volume=

    Reinforcement learning for solving the vehicle routing problem , author=. Advances in Neural Information Processing Systems , volume=

  44. [44]

    Solving corrective risk-based security-constrained optimal power flow with

    Wang, Qin and McCalley, James D and Zheng, Tongxin and Litvinov, Eugene , journal=. Solving corrective risk-based security-constrained optimal power flow with. 2016 , publisher=

    2016

  45. [45]

    An accelerated

    Pishvaee, Mir Saman and Razmi, Jafar and Torabi, Seyed Ali , journal=. An accelerated. 2014 , publisher=

    2014

  46. [46]

    Maravelias, Christos T and Grossmann, Ignacio E , booktitle=. Using. 2004 , organization=

    2004

  47. [47]

    Application of

    Canto, Salvador Perez , journal=. Application of. 2008 , publisher=

    2008

  48. [48]

    Benders Decomposition for stochastic programming-based

    Alhaider, Mohemmed and Fan, Lingling and Miao, Zhixin , booktitle=. Benders Decomposition for stochastic programming-based. 2016 , organization=

    2016

  49. [49]

    A hybrid outer-approximation/

    De Camargo, Ricardo Saraiva and de Miranda Jr, Gilberto and Ferreira, Ricardo PM , journal=. A hybrid outer-approximation/. 2011 , publisher=

    2011

  50. [50]

    Solving large nonconvex water resources management models using generalized

    Cai, Ximing and McKinney, Daene C and Lasdon, Leon S and Watkins Jr, David W , journal=. Solving large nonconvex water resources management models using generalized. 2001 , publisher=

    2001

  51. [51]

    Decomposition based hybrid

    Behnamian, Javad , journal=. Decomposition based hybrid. 2014 , publisher=

    2014

  52. [52]

    A nested

    Naoum-Sawaya, Joe and Elhedhli, Samir , journal=. A nested. 2010 , publisher=

    2010

  53. [53]

    Crainic, Teodor Gabriel and Rei, Walter and Hewitt, Mike and Maggioni, Francesca , volume=. Partial. 2016 , publisher=

    2016

  54. [54]

    A branch-and-

    Gendron, Bernard and Scutell. A branch-and-. European Journal of Operational Research , volume=. 2016 , publisher=

    2016

  55. [55]

    Accelerating

    Oliveira, Fabricio and Grossmann, Ignacio E and Hamacher, Silvio , journal=. Accelerating. 2014 , publisher=

    2014

  56. [56]

    Mathematical Programming , volume=

    A cross-decomposition scheme with integrated primal--dual multi-cuts for two-stage stochastic programming investment planning problems , author=. Mathematical Programming , volume=. 2016 , publisher=

    2016

  57. [57]

    Practical enhancements to the Magnanti--

    Papadakos, Nikolaos , journal=. Practical enhancements to the Magnanti--. 2008 , publisher=

    2008

  58. [58]

    On generating maximal nondominated

    Sherali, Hanif D and Lunday, Brian J , journal=. On generating maximal nondominated. 2013 , publisher=

    2013

  59. [59]

    Multicommodity distribution system design by

    Geoffrion, Arthur M and Graves, Glenn W , journal=. Multicommodity distribution system design by. 1974 , publisher=

    1974

  60. [60]

    IEEE/ACM Transactions on Networking , volume=

    Exact and heuristic algorithms for data-gathering cluster-based wireless sensor network design problem , author=. IEEE/ACM Transactions on Networking , volume=. 2013 , publisher=

    2013

  61. [61]

    Telecommunication Systems , volume=

    Benders decomposition algorithms for the fixed-charge relay network design in telecommunications , author=. Telecommunication Systems , volume=. 2014 , publisher=

    2014

  62. [62]

    European Journal of Operational Research , volume=

    Decomposition based hybrid metaheuristics , author=. European Journal of Operational Research , volume=. 2015 , publisher=

    2015

  63. [63]

    Speeding up logic-based

    Raidl, G. Speeding up logic-based. International Workshop on Hybrid Metaheuristics , pages=. 2014 , organization=

    2014

  64. [64]

    Improving

    Poojari, Chandra A and Beasley, John E , journal=. Improving. 2009 , publisher=

    2009

  65. [65]

    A hybrid

    Sohn, Hansuk and Tseng, Bill and Bricker, Dennis L , journal=. A hybrid

  66. [66]

    A hybrid algorithm of tabu search and

    Jiang, Wei and Tang, Lixin and Xue, Shixu , booktitle=. A hybrid algorithm of tabu search and. 2009 , organization=

    2009

  67. [67]

    2019 American Control Conference (ACC) , pages=

    Safe and near-optimal policy learning for model predictive control using primal-dual neural networks , author=. 2019 American Control Conference (ACC) , pages=. 2019 , organization=

    2019

  68. [68]

    Automatica , volume=

    Large scale model predictive control with neural networks and primal active sets , author=. Automatica , volume=. 2022 , publisher=

    2022

  69. [69]

    Multi-agent sensitivity enhanced iterative best response:

    Wang, Zijian and Taubner, Tim and Schwager, Mac , journal=. Multi-agent sensitivity enhanced iterative best response:. 2020 , publisher=

    2020

  70. [70]

    IFAC-PapersOnLine , volume=

    Fast multi-robot motion planning via imitation learning of mixed-integer programs , author=. IFAC-PapersOnLine , volume=. 2021 , publisher=

    2021

  71. [71]

    2019 18th European Control Conference (ECC) , pages=

    Learning binary warm starts for multiparametric mixed-integer quadratic programming , author=. 2019 18th European Control Conference (ECC) , pages=. 2019 , organization=

    2019

  72. [72]

    INFORMS Journal on Computing , volume=

    Online mixed-integer optimization in milliseconds , author=. INFORMS Journal on Computing , volume=. 2022 , publisher=

    2022

  73. [73]

    Learning to initialize generalized

    Mitrai, Ilias and Daoutidis, Prodromos , journal=. Learning to initialize generalized

  74. [74]

    Taking the human out of decomposition-based optimization via artificial intelligence,

    Mitrai, Ilias and Daoutidis, Prodromos , journal=. Taking the human out of decomposition-based optimization via artificial intelligence,. 2024 , publisher=

    2024

  75. [75]

    Accelerating generalized

    Lee, Mengyuan and Ma, Ning and Yu, Guanding and Dai, Huaiyu , journal=. Accelerating generalized. 2020 , publisher=

    2020

  76. [76]

    Machine learning-empowered

    Wu, Tao and Chen, Weiwei and Cordeau, Jean-Fran. Machine learning-empowered. INFORMS Journal on Computing , year=

  77. [77]

    Towards accelerating

    Mak, Stephen and Mana, Kyle and Zehtabi, Parisa and Cashmore, Michael and Magazzeni, Daniele and Veloso, Manuela , booktitle=. Towards accelerating

  78. [78]

    Reinforcement learning with combinatorial actions:

    Delarue, Arthur and Anderson, Ross and Tjandraatmadja, Christian , journal=. Reinforcement learning with combinatorial actions:

  79. [79]

    Advances in Neural Information Processing Systems , volume=

    Exact combinatorial optimization with graph convolutional neural networks , author=. Advances in Neural Information Processing Systems , volume=

  80. [80]

    Solving mixed integer programs using neural networks

    Solving mixed integer programs using neural networks , author=. arXiv preprint arXiv:2012.13349 , year=

    work page arXiv 2012

Showing first 80 references.