Pith. sign in

REVIEW 3 major objections 4 minor 50 references

On Distributional Dependent Performance of Classical and Neural Routing Solvers

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Neural CVRP solvers beat HGS when trained on a fixed node pool

desk verdict Novel subsampling protocol for NCO, but the headline 'beats HGS' result is not established because test nodes come from the training pool; the gap-narrowing claim is plausible. read the letter →

arxiv 2508.02510 v1 pith:742PR4AS submitted 2025-08-04 cs.LG

Daniela Thyssens , Tim Dernedde , Wilson Sentanoe , Lars Schmidt-Thieme This is my paper · ORCID
classification cs.LG
keywords neuralcombinatorialoptimizationroutingproblemscapacitatedvehicleproblemtravelingsalesmansubsamplingbasenodedistributiondistributionalalignmentmeta-heuristics
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

This paper aims to establish that neural routing solvers can close—and in some cases reverse—the performance gap to handcrafted optimization heuristics when the training data is subsampled from a fixed base node distribution. Instead of drawing each instance's customers independently, the authors build one large pool of nodes for each problem class and draw every training and test instance as a uniform subset of that same pool, so customers recur across instances. On the Uchoa X-based CVRP benchmarks they report negative relative gaps against the leading OR solver HGS under equal wall-clock budgets, with the largest gap at $-4.22\%$ for the SGBS search method under a 50-second budget. If true, this means distributional alignment is a first-order lever for solver performance, not just network architecture. The paper also shows the effect is problem-dependent: on TSP, classical solvers remain ahead.

What carries the argument

The load-bearing object is the Base Node Distribution: a fixed pool of $N_{base}$ customer nodes (coordinates, demands, and a fixed depot for CVRP) drawn once from a designated underlying distribution such as Uniform, Rotation, Explosion, or Uchoa's random-clustered X family. The subsampling function $f_{sample}$ produces each training and test instance as a uniform i.i.d. subset of that pool, which plants recurring nodes across instances and gives the learned policy a stable structure to exploit. The measuring protocol matters just as much as the data generation: every solver runs under the same per-instance wall-clock budget $T_{MAX}$, and performance is reported as the percentage gap to LKH3 (for TSP) or HGS (for CVRP). That equal-budget comparison is what lets a negative gap count as an outright win for the neural method.

What would settle it

Train the same neural methods on subsamples of one base node pool and evaluate them on subsamples of a second, held-out pool generated from the same underlying distribution with the same node counts and runtime budgets; if the negative relative gaps to HGS vanish or turn positive, node-level memorization of the training pool rather than generalizable routing skill is producing the paper's headline result.

Watch

Extended reading notes

Core claim

The paper's central claim is that a subsampling protocol, formalized in Eqs. (2)--(4), turns learned routing solvers into strong competitors for classical OR methods. A base node distribution $G^j_{base}$ is sampled once from an underlying distribution $\mathcal{D}_j$; each instance is a uniform subsample of those $N_{base}$ nodes, with a fixed depot added for the CVRP. Because nodes reappear across training and test instances, the model can internalize the pool's spatial and demand structure. Under equal per-instance runtime budgets, subsampled versions of POMO, SGBS, and NeuOpt obtain negative relative gaps to HGS on the Uchoa X (random-clustered) CVRP data, for instance SGBS at $-4.22\%$ at 50 seconds, and BQ reaches $-0.01\%$ on the TSP X200 set at 0.7 seconds. The authors read these numbers as evidence that neural solvers are viable, adaptive alternatives when the test distribution matches the distribution they were trained on.

Load-bearing premise

The load-bearing premise is that a test set subsampled from the same base node pool used in training is a fair measure of solver quality, even though those exact nodes appeared in training instances for the neural models and are unseen for the classical solvers.

Editorial extensions

If this is right

  • If the reported negative gaps hold, learned CVRP solvers can serve as primary solvers in structured, distribution-matched logistics settings rather than merely as initializers for classical heuristics.
  • Evaluation protocols for neural combinatorial optimization should include subsampled, equal-time-budget benchmarks; the standard instance-by-instance i.i.d. sampling appears to understate what neural solvers achieve on structured distributions.
  • Distributional alignment becomes an explicit, tunable design axis for routing solvers, alongside architecture and training objective.
  • The same protocol does not close the gap on TSP, so the benefit is problem-dependent and not a general repeal of classical solver dominance.
  • Recurring customers from a fixed pool mirror real logistics networks where the same clients are served in varying daily configurations, which gives the performance gain a direct practical interpretation.

Reading between the lines

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

  • The paper leaves open whether node-level memorization contributes to the negative gaps: every test node appeared in training instances, while HGS and LKH3 see those nodes only at test time, so the reported result does not by itself measure generalizable routing skill.
  • A natural test the authors do not run is to train on one base pool and evaluate on a held-out pool drawn from the same underlying distribution; if the negative gaps shrink or vanish, the advantage is partly memorization of the coordinate and demand pool rather than learning the distribution itself.
  • The subsampling idea should transfer to other combinatorial problems with recurring ground elements, such as scheduling with recurring jobs or warehouse picking with recurring SKUs, where a fixed pool is a natural inductive bias.
  • Mixing full-distribution and subsampled training, or gradually enlarging the base pool during training, could combine the specialization reported here with the robustness classical solvers retain across diverse distributions; the paper lists hybrid training as future work.
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

3 major / 4 minor

Summary. This paper proposes training neural routing solvers (POMO, BQ, SGBS-EAS, NeuOpt) on subsamples drawn from a fixed 'base node distribution' rather than on independently sampled instances, with the motivation that recurring customers reflect practical logistics. The authors evaluate the trained models on subsampled test sets for TSP and CVRP and compare them against LKH3 and HGS-CVRP under time budgets of 0.7s, 5s, and 50s. They report that subsampling reduces the gap relative to full-distribution training and, in Table 4b, report negative gaps on the CVRP Uchoa-style GX10k test set, which they interpret as neural solvers outperforming state-of-the-art OR meta-heuristics. The paper also discusses differences across four underlying distributions.

Significance. The idea of testing neural routing solvers under a repeated-customer protocol is practically motivated and the paper is clearly written. If the negative gaps in Table 4b persisted when the test base distribution is disjoint from the training base distribution, the result would be a significant step for distribution-aware NCO. However, the current evaluation allows node-level overlap between training and test sets, so the headline comparisons with HGS and LKH3 conflate distributional learning with memorization of the fixed node pool. The contribution is therefore not yet established as stated; the weaker claim that the gap shrinks in a repeated-customer setting may be defensible, but the stronger claim that neural solvers surpass OR solvers needs a corrected protocol.

major comments (3)
  1. [Section 3.2, Eq. (4); Table 4b] The test instances G^j_test are subsampled from the same G^j_base used to generate G^j_train, so the test nodes have appeared in training. With Nbase=200 every test node is in the training pool; with Nbase=10000 each test node is seen many times during on-the-fly training. The OR baselines see those nodes only at test time. The negative gaps on GX10k in Table 4b are therefore compatible with node-pool memorization (coordinate/demand lookup and pool-specific route fragments) rather than with a general ability to beat HGS. This is the main load-bearing issue for the Section 5 claim that subsampled models 'outperformed state-of-the-art OR solvers.' Please add an experiment with a held-out base node distribution (freshly sampled from the same underlying Dj) and report whether negative gaps survive; if they do not, the claim should be restricted to the repeated-customer setting, and Table 3 should be interpreted with the same caveat because the full-distribution baselines have not seen the test nodes at all.
  2. [Appendix D; Tables 3 and 4] Appendix D states that 'we evaluated all models once' on each dataset. Many gaps in Tables 3 and 4 are within one or two percentage points, and single runs provide no information about variance. Without repeated runs or per-instance standard errors, the reported negative gaps (e.g., SGBS -4.22% on GX10k at TMAX=50) cannot be distinguished from run-to-run noise. Please provide multiple seeds/runs and confidence intervals, or at least the per-instance distribution of gaps.
  3. [Table 4b, GX10k row at TMAX=0.7] The entries for POMOsub and SGBSsub are identical (-2.564), which is either a typographical error or a data-processing artifact. Because this row is part of the evidence for the central claim, please verify all values in Table 4b and correct or explain the coincidence.
minor comments (4)
  1. [Section 3.2, Eq. (3)] The statement 'G^i_j i.i.d. ∼ G^j_base' is not precise: a subsample drawn without replacement from a fixed base set is not an i.i.d. draw from the underlying instance distribution. Please state whether subsampling is with or without replacement and clarify the sampling distribution.
  2. [Section 4.1 / Data availability] The contribution list promises new datasets and data generators, but the manuscript does not provide a code/data link or supplementary material. Please include the repository URL.
  3. [Table 3] The table columns are not visually separated in the current manuscript; entries such as '1.2190 1.646 1.675 4.598 5.848 4.559 2.702 1.354' appear as a single block, making it hard to read. Please format as a proper table.
  4. [Section 5] The language 'neural solvers are not just scalable, but capable of matching and exceeding handcrafted algorithms' goes beyond the results, which are limited to the repeated-customer protocol; please qualify this statement to match the evidence.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline negative gaps against HGS/LKH are measured on test sets drawn from the same base node pool used for training, so the claimed outperformance is an artifact of the split by construction.

  1. self definitional [Section 3.2, Eqs. (3)-(4); Section 4.3, Table 4b; Section 5]
    "Finally we define the train and test sets for an underlying distribution Dj and a respective base node distribution Gj base as follows: Gj train := {Gj i }Ltrain i=1 , Gj test := (Gj i )Ltrain+Ltest i=Ltrain+1 (4) ... The train and test instances are subsampled with different seeds respectively to avoid training on test. ... Most notably, for CVRP on the Uchoa benchmark, subsampled models outperformed state-of-the-art OR solvers (HGS) under realistic runtime constraints, achieving negative relative gaps."

    Eq. (4) defines the training and test sets as subsamples of the same Gj_base via the same fsample. A 'different seed' prevents identical instance copies, but not node reuse; for Nbase=200/500 with N=100, every test node has appeared in training instances, and even for Nbase=10,000 the RL training on the fly visits the fixed pool repeatedly. The negative gaps in Table 4b (e.g., SGBS -4.22% on GX10k at TMAX=50s) are therefore measured against LKH3/HGS, which see each test instance from scratch, while the neural policy has been trained on the exact coordinate/demand pool.

full rationale

The paper is an empirical study rather than a formal derivation, so the classical fitted-parameter or self-citation circularities are absent; the same-group citations ([1], [12], [13]) are not load-bearing. The circularity is in the evaluation of the central claim. Equations (2)-(4) make both training and test sets subsamples of the same base node distribution Gj_base, and 'different seeds' only avoids duplicate instances, not node overlap. Consequently, Table 4b's negative relative gaps on the X/Uchoa CVRP benchmarks compare pool-familiar neural policies against OR solvers that see the test nodes for the first time. The paper's weaker claim that subsampling narrows the gap retains some support from the full-distribution controls in Table 3, and many configurations still show positive gaps, so the circularity is partial rather than total. Appendix D's admission that all models were evaluated once is a statistical limitation and further weakens the comparison, but it is not itself a circular step.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper is an empirical study, not a derivation, so the ledger records hand-chosen experimental design choices and assumptions. The most consequential item is the shared base node pool for train and test, which is an evaluation assumption rather than a fitted parameter.

free parameters (2)
  • Base node distribution size Nbase = 200, 500, and 10000
    Hand-chosen pool sizes appear in Table 1; they control the degree of node re-occurrence and therefore the magnitude of the reported gap reductions.
  • Runtime budget TMAX = 0.7, 5, and 50 seconds
    Hand-chosen time limits in Section 4.3; the negative gaps against HGS appear only under these specific budgets.
assumptions (4)
  • domain assumption Train and test instances are subsampled from the same fixed base node distribution, so test nodes co-occur with training nodes.
    Section 3.2, Eqs. (2)-(4). This is the core evaluation assumption and it enables node memorization.
  • domain assumption Node re-occurrence across instances is a valid model of real-world logistics and a legitimate basis for benchmarking learned solvers.
    Section 1, 'regular retail clients' argument. Without this, the protocol is a train/test leakage rather than a realistic scenario.
  • domain assumption Equal wall-clock budget TMAX gives a fair comparison between neural solvers and OR metaheuristics.
    Section 4.1, metrics paragraph. The budget does not account for the offline training the neural methods receive on the test node pool.
  • ad hoc to paper A single run per configuration is sufficient to draw conclusions.
    Appendix D states all models were evaluated once, which 'could potentially affect statistical significance.'

how reviews work

0 comments
Cite this review

Pith. "Pith review of On Distributional Dependent Performance of Classical and Neural Routing Solvers." pith.science (2026). https://pith.science/paper/742PR4AS

@misc{pith2026250802510,
  author       = {Pith},
  title        = {Pith review of: On Distributional Dependent Performance of Classical and Neural Routing Solvers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/742PR4AS}},
  note         = {Machine review of arXiv:2508.02510}
}
read the original abstract

Neural Combinatorial Optimization aims to learn to solve a class of combinatorial problems through data-driven methods and notably through employing neural networks by learning the underlying distribution of problem instances. While, so far neural methods struggle to outperform highly engineered problem specific meta-heuristics, this work explores a novel approach to formulate the distribution of problem instances to learn from and, more importantly, plant a structure in the sampled problem instances. In application to routing problems, we generate large problem instances that represent custom base problem instance distributions from which training instances are sampled. The test instances to evaluate the methods on the routing task consist of unseen problems sampled from the underlying large problem instance. We evaluate representative NCO methods and specialized Operation Research meta heuristics on this novel task and demonstrate that the performance gap between neural routing solvers and highly specialized meta-heuristics decreases when learning from sub-samples drawn from a fixed base node distribution.

Figures

Figures reproduced from arXiv: 2508.02510 by the authors.

Figure 1
Figure 1. Overview Subsampling Approach. The Base Node Distribution [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Base Node Distributions G j base (top) from which routing problems Gi are subsam￾pled (bottom). NCO methods have employed various sampling strategies for training on routing tasks. Notably, RL￾based approaches are primarily trained on data generated during each epoch of training [36, 25, 34]. To efficiently incorporate subsampling into the training of different NCO methods, we define a generic NCO Dataset class, whi… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 41 canonical work pages

  1. [1]

    Attention, filling in the gaps for generalization in routing problems

    Ahmad Bdeir, Jonas K Falkner, and Lars Schmidt-Thieme. Attention, filling in the gaps for generalization in routing problems. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 505–520. Springer, 2022

    work page 2022

  2. [2]

    Le, Mohammad Norouzi, and Samy Bengio

    Irwan Bello, Hieu Pham, Quoc V . Le, Mohammad Norouzi, and Samy Bengio. Neural combi- natorial optimization with reinforcement learning. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Workshop Track Proceedings. OpenReview.net, 2017. URL https://openreview.net/forum?id=Bk9mxlSFx

    work page 2017

  3. [3]

    Machine learning for combinatorial optimization: A methodological tour d’horizon

    Yoshua Bengio, Andrea Lodi, and Antoine Prouvost. Machine learning for combinatorial optimization: A methodological tour d’horizon. Eur. J. Oper. Res., 290(2):405–421, 2021. doi: 10.1016/j.ejor.2020.07.063. URL https://doi.org/10.1016/j.ejor.2020.07.063

    doi:10.1016/j.ejor.2020.07.063 2021

  4. [4]

    RouteFinder: Towards Foundation Models for Vehicle Routing Problems, June 2024

    Federico Berto, Chuanbo Hua, Nayeli Gast Zepeda, André Hottung, Niels Wouda, Leon Lan, Kevin Tierney, and Jinkyoo Park. RouteFinder: Towards Foundation Models for Vehicle Routing Problems, June 2024

    work page 2024

  5. [5]

    Learning Generalizable Models for Vehicle Routing Problems via Knowledge Distillation

    Jieyi Bi, Yining Ma, Jiahai Wang, Zhiguang Cao, Jinbiao Chen, Yuan Sun, and Yeow Meng Chee. Learning generalizable models for vehicle routing problems via knowledge distillation. arXiv preprint arXiv:2210.07686, 2022

    work page Pith review arXiv 2022

  6. [6]

    Evolving diverse tsp instances by means of novel and creative mutation operators

    Jakob Bossek, Pascal Kerschke, Aneta Neumann, Markus Wagner, Frank Neumann, and Heike Trautmann. Evolving diverse tsp instances by means of novel and creative mutation operators. In Proceedings of the 15th ACM/SIGEVO conference on foundations of genetic algorithms , pages 58–71, 2019

    work page 2019

  7. [7]

    The Transformer Network for the Traveling Salesman Problem, March 2021

    Xavier Bresson and Thomas Laurent. The Transformer Network for the Traveling Salesman Problem, March 2021

    work page 2021

  8. [8]

    Combinatorial Optimization with Policy Adaptation using Latent Space Search

    Felix Chalumeau, Shikha Surana, Clément Bonnet, Nathan Grinsztajn, Arnu Pretorius, Alexan- dre Laterre, and Tom Barrett. Combinatorial Optimization with Policy Adaptation using Latent Space Search. In Advances in Neural Information Processing Systems , volume 36, pages 7947–7959, December 2023

    work page 2023

Show all 50 references
  1. [9]

    Simulation-guided Beam Search for Neural Combinatorial Optimization

    Jinho Choo, Yeong-Dae Kwon, Jihoon Kim, Jeongwoo Jae, André Hottung, Kevin Tierney, and Youngjune Gwon. Simulation-guided Beam Search for Neural Combinatorial Optimization. In Sanmi Koyejo, S. Mohamed, A. Agarwal, Danielle Belgrave, K. Cho, and A. Oh, editors, Advances in Neur...

    2022

  2. [10]

    Learning Heuristics for the TSP by Policy Gradient

    Michel Deudon, Pierre Cournut, Alexandre Lacoste, Yossiri Adulyasak, and Louis-Martin Rousseau. Learning Heuristics for the TSP by Policy Gradient. In Willem-Jan van Hoeve, editor, Integration of Constraint Programming, Artificial Intelligence, and Operations Research, Lecture...

    2018 doi

  3. [11]

    BQ-NCO: Bisimulation Quotienting for Efficient Neural Combinatorial Optimization

    Darko Drakulic, Sofia Michel, Florian Mai, Arnaud Sors, and Jean-Marc Andreoli. BQ-NCO: Bisimulation Quotienting for Efficient Neural Combinatorial Optimization. In Thirty-Seventh Conference on Neural Information Processing Systems, November 2023

    2023

  4. [12]

    Falkner and Lars Schmidt-Thieme

    Jonas K. Falkner and Lars Schmidt-Thieme. Too Big, so Fail? - Enabling Neural Construction Methods to Solve Large-Scale Routing Problems. CoRR, abs/2309.17089, 2023. doi: 10.48550/ ARXIV .2309.17089

    2023 doi arXiv

  5. [13]

    Learning to control local search for combinatorial optimization

    Jonas K Falkner, Daniela Thyssens, Ahmad Bdeir, and Lars Schmidt-Thieme. Learning to control local search for combinatorial optimization. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 361–376. Springer, 2022. 10

    2022

  6. [14]

    Generalize a Small Pre-trained Model to Arbitrarily Large TSP Instances

    Zhang-Hua Fu, Kai-Bin Qiu, and Hongyuan Zha. Generalize a Small Pre-trained Model to Arbitrarily Large TSP Instances. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021...

    2021 doi

  7. [15]

    Generalization of neural combinatorial solvers through the lens of adversarial robustness

    Simon Geisler, Johanna Sommer, Jan Schuchardt, Aleksandar Bojchevski, and Stephan Günne- mann. Generalization of neural combinatorial solvers through the lens of adversarial robustness. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event...

    2022

  8. [16]

    An extension of the lin-kernighan-helsgaun tsp solver for constrained traveling salesman and vehicle routing problems

    Keld Helsgaun. An extension of the lin-kernighan-helsgaun tsp solver for constrained traveling salesman and vehicle routing problems. Technical Report, Roskilde: Roskilde University, 12, 2017

    2017

  9. [17]

    Neural large neighborhood search for the capacitated vehicle routing problem

    André Hottung and Kevin Tierney. Neural large neighborhood search for the capacitated vehicle routing problem. In Giuseppe De Giacomo, Alejandro Catalá, Bistra Dilkina, Michela Milano, Senén Barro, Alberto Bugarín, and Jérôme Lang, editors, ECAI 2020 - 24th European Conference...

    2020 doi

  10. [18]

    Neural large neighborhood search for routing problems

    André Hottung and Kevin Tierney. Neural large neighborhood search for routing problems. Artif. Intell., 313:103786, 2022. doi: 10.1016/J.ARTINT.2022.103786

    2022

  11. [19]

    Efficient Active Search for Combinatorial Optimization Problems

    André Hottung, Yeong-Dae Kwon, and Kevin Tierney. Efficient Active Search for Combinatorial Optimization Problems. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022

    2022

  12. [20]

    Efficient active search for combinatorial optimization problems

    André Hottung, Yeong-Dae Kwon, and Kevin Tierney. Efficient active search for combinatorial optimization problems. In The Tenth International Conference on Learning Representations, ICLR. OpenReview.net, 2022. URL https://openreview.net/forum?id=nO5caZwFwYu

    2022

  13. [21]

    PolyNet: Learning Diverse Solution Strategies for Neural Combinatorial Optimization, February 2024

    André Hottung, Mridul Mahajan, and Kevin Tierney. PolyNet: Learning Diverse Solution Strategies for Neural Combinatorial Optimization, February 2024

    2024

  14. [22]

    Pointerformer: Deep reinforced multi-pointer transformer for the traveling salesman problem

    Yan Jin, Yuandong Ding, Xuanhao Pan, Kun He, Li Zhao, Tao Qin, Lei Song, and Jiang Bian. Pointerformer: Deep reinforced multi-pointer transformer for the traveling salesman problem. In Brian Williams, Yiling Chen, and Jennifer Neville, editors, Thirty-Seventh AAAI Conference o...

    2023

  15. [23]

    Sym-NCO: Leveraging Symmetricity for Neural Combinatorial Optimization

    Minsu Kim, Junyoung Park, and Jinkyoo Park. Sym-NCO: Leveraging Symmetricity for Neural Combinatorial Optimization. In Advances in Neural Information Processing Systems, October 2022

    2022

  16. [24]

    Scale-conditioned adaptation for large scale combinatorial optimization

    Minsu Kim, Jiwoo Son, Hyeonah Kim, and Jinkyoo Park. Scale-conditioned adaptation for large scale combinatorial optimization. In NeurIPS 2022 Workshop on Distribution Shifts: Connecting Methods and Applications, 2022

    2022

  17. [25]

    Attention, Learn to Solve Routing Problems! In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019

    Wouter Kool, Herke van Hoof, and Max Welling. Attention, Learn to Solve Routing Problems! In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019

    2019

  18. [26]

    Deep policy dynamic programming for vehicle routing problems

    Wouter Kool, Herke van Hoof, Joaquim Gromicho, and Max Welling. Deep policy dynamic programming for vehicle routing problems. In Integration of Constraint Programming, Artificial Intelligence, and Operations Research: 19th International Conference, CPAIOR 2022, Los Angeles, CA...

    2022

  19. [27]

    Wouter Kool, Herke van Hoof, Joaquim A. S. Gromicho, and Max Welling. Deep Policy Dynamic Programming for Vehicle Routing Problems. In Pierre Schaus, editor, Integration of Constraint Programming, Artificial Intelligence, and Operations Research - 19th International Conference...

    2022

  20. [28]

    POMO: Policy optimization with multiple optima for reinforcement learning

    Yeong-Dae Kwon, Jinho Choo, Byoungjip Kim, Iljoo Yoon, Youngjune Gwon, and Seungjai Min. POMO: Policy optimization with multiple optima for reinforcement learning. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Advanc...

    2020

  21. [29]

    Matrix encoding networks for neural combinatorial optimization

    Yeong-Dae Kwon, Jinho Choo, Iljoo Yoon, Minah Park, Duwon Park, and Youngjune Gwon. Matrix encoding networks for neural combinatorial optimization. Advances in Neural Informa- tion Processing Systems, 34:5138–5149, 2021

    2021

  22. [30]

    Learning to delegate for large-scale vehicle routing

    Sirui Li, Zhongxia Yan, and Cathy Wu. Learning to delegate for large-scale vehicle routing. In Marc’Aurelio Ranzato, Alina Beygelzimer, Yann N. Dauphin, Percy Liang, and Jennifer Wort- man Vaughan, editors, Advances in Neural Information Processing Systems 34: Annual Confer- e...

    2021

  23. [31]

    From Distribution Learning in Training to Gradient Search in Testing for Combinatorial Optimization

    Yang Li, Jinpei Guo, Runzhong Wang, and Junchi Yan. From Distribution Learning in Training to Gradient Search in Testing for Combinatorial Optimization. In Thirty-Seventh Conference on Neural Information Processing Systems, November 2023

    2023

  24. [32]

    Evaluating curriculum learning strategies in neural combinatorial optimization

    Michal Lisicki, Arash Afkanpour, and Graham W Taylor. Evaluating curriculum learning strategies in neural combinatorial optimization. arXiv preprint arXiv:2011.06188, 2020

    2011 arXiv

  25. [33]

    Neural Combinatorial Optimization with Heavy Decoder: Toward Large Scale Generalization

    Fu Luo, Xi Lin, Fei Liu, Qingfu Zhang, and Zhenkun Wang. Neural Combinatorial Optimization with Heavy Decoder: Toward Large Scale Generalization. In Thirty-Seventh Conference on Neural Information Processing Systems, November 2023

    2023

  26. [34]

    Learning to Search Feasible and Infeasible Regions of Routing Problems with Flexible Neural k-Opt

    Yining Ma, Zhiguang Cao, and Yeow Meng Chee. Learning to Search Feasible and Infeasible Regions of Routing Problems with Flexible Neural k-Opt. In Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine, editors, Advances in Neural Information P...

    2023

  27. [35]

    Yimeng Min, Yiwei Bai, and Carla P. Gomes. Unsupervised Learning for Solving the Travelling Salesman Problem. In Thirty-Seventh Conference on Neural Information Processing Systems, November 2023

    2023

  28. [36]

    Reinforce- ment learning for solving the vehicle routing problem

    MohammadReza Nazari, Afshin Oroojlooy, Lawrence Snyder, and Martin Takac. Reinforce- ment learning for solving the vehicle routing problem. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Pro- cessin...

    2018

  29. [37]

    DIFUSCO: Graph-based Diffusion Solvers for Combinato- rial Optimization

    Zhiqing Sun and Yiming Yang. DIFUSCO: Graph-based Diffusion Solvers for Combinato- rial Optimization. In Thirty-Seventh Conference on Neural Information Processing Systems, November 2023

    2023

  30. [38]

    New benchmark instances for the capacitated vehicle routing problem

    Eduardo Uchoa, Diego Pecin, Artur Pessoa, Marcus Poggi, Thibaut Vidal, and Anand Subrama- nian. New benchmark instances for the capacitated vehicle routing problem. European Journal of Operational Research, 257(3):845–858, 2017

    2017

  31. [39]

    Hybrid genetic search for the cvrp: Open-source implementation and swap* neighborhood

    Thibaut Vidal. Hybrid genetic search for the cvrp: Open-source implementation and swap* neighborhood. Computers & Operations Research, 140:105643, 2022

    2022

  32. [40]

    A hybrid genetic algorithm for multidepot and periodic vehicle routing problems

    Thibaut Vidal, Teodor Gabriel Crainic, Michel Gendreau, Nadia Lahrichi, and Walter Rei. A hybrid genetic algorithm for multidepot and periodic vehicle routing problems. Operations Research, 60(3):611–624, 2012. 12

    2012

  33. [41]

    Pointer Networks

    Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer Networks. In NIPS, 2015

    2015

  34. [42]

    Multi-Decoder Attention Model with Em- bedding Glimpse for Solving Vehicle Routing Problems

    Liang Xin, Wen Song, Zhiguang Cao, and Jie Zhang. Multi-Decoder Attention Model with Em- bedding Glimpse for Solving Vehicle Routing Problems. Proceedings of the AAAI Conference on Artificial Intelligence, 35(13):12042–12049, May 2021. ISSN 2374-3468, 2159-5399. doi: 10.1609/a...

    2021 doi

  35. [43]

    NeuroLKH: Combining Deep Learning Model with Lin-Kernighan-Helsgaun Heuristic for Solving the Traveling Salesman Problem

    Liang Xin, Wen Song, Zhiguang Cao, and Jie Zhang. NeuroLKH: Combining Deep Learning Model with Lin-Kernighan-Helsgaun Heuristic for Solving the Traveling Salesman Problem. In Marc’Aurelio Ranzato, Alina Beygelzimer, Yann N. Dauphin, Percy Liang, and Jennifer Wort- man Vaughan,...

    2021

  36. [44]

    DeepACO: Neural-enhanced Ant Systems for Combinatorial Optimization

    Haoran Ye, Jiarui Wang, Zhiguang Cao, Helan Liang, and Yong Li. DeepACO: Neural-enhanced Ant Systems for Combinatorial Optimization. In DeepACO: Neural-enhanced Ant Systems for Combinatorial Optimization. arXiv, September 2023

    2023

  37. [45]

    GLOP: Learning Global Partition and Local Construction for Solving Large-Scale Routing Problems in Real-Time

    Haoran Ye, Jiarui Wang, Helan Liang, Zhiguang Cao, Yong Li, and Fanzhang Li. GLOP: Learning Global Partition and Local Construction for Solving Large-Scale Routing Problems in Real-Time. In Michael J. Wooldridge, Jennifer G. Dy, and Sriraam Natarajan, editors, Thirty-Eighth AA...

    2024

  38. [46]

    It’s Not What Machines Can Learn, It’s What We Cannot Teach

    Gal Yehuda, Moshe Gabel, and Assaf Schuster. It’s Not What Machines Can Learn, It’s What We Cannot Teach. In Proceedings of the 37th International Conference on Machine Learning, pages 10831–10841. PMLR, November 2020

    2020

  39. [47]

    Learning to solve travelling salesman problem with hardness-adaptive curriculum

    Zeyang Zhang, Ziwei Zhang, Xin Wang, and Wenwu Zhu. Learning to solve travelling salesman problem with hardness-adaptive curriculum. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intel- ...

    2022 doi

  40. [48]

    UDC: A Unified Neural Divide-and-Conquer Framework for Large-Scale Combinatorial Optimization Problems, June 2024

    Zhi Zheng, Changliang Zhou, Tong Xialiang, Mingxuan Yuan, and Zhenkun Wang. UDC: A Unified Neural Divide-and-Conquer Framework for Large-Scale Combinatorial Optimization Problems, June 2024

    2024

  41. [49]

    Towards omni-generalizable neural methods for vehicle routing problems

    Jianan Zhou, Yaoxin Wu, Wen Song, Zhiguang Cao, and Jie Zhang. Towards omni-generalizable neural methods for vehicle routing problems. In International Conference on Machine Learning, pages 42769–42789. PMLR, 2023. 13 A Routing Problem Definitions The TSP. The TSP consists of ...

    2023

  42. [50]

    here. POMO Policy Optimization with Multiple Optima [28] proposed a training and inference mecha- nism for constructive models where they adjust the baseline function in the policy gradient, averaging over multiple rollouts with different starting nodes for a problem instance ...

    2012

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.