Pith. sign in

REVIEW 4 major objections 5 minor 77 references

Graph Neural Network-based Algorithm Selection for the Traveling Salesman Problem: A Systematic Study of Cost and Rank Losses under Distinct Budget Regimes

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a graph neural network reading raw TSP instances can pick, per instance, the best of five solvers and improve normalized solution cost over the best fixed solver under both 10- and 60-second budgets.

desk verdict A clean empirical study of GNN-based TSP algorithm selection, but the headline significance claims rest on the same 300 instances used to pick the loss function; the descriptive improvements are real, the p-values are not. read the letter →

arxiv 2607.18632 v1 pith:4LHBOAQH submitted 2026-07-21 cs.LG

Zhaoxuan Li , Jiale Yang , Yifei Lu , Mustafa Misir This is my paper
classification cs.LG MSC 68T0790C27
keywords algorithmselectiontravelingsalesmanproblemgraphneuralnetworkslearningtorankcostpredictionlossfunctionsruntimebudgetssolverportfolio
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

GNNAS-TSP is an algorithm-selection system for the traveling salesman problem: it takes a TSP instance as a raw graph, predicts what each of five solvers would cost on that instance under a fixed time budget, and runs the solver predicted to be cheapest. The central claim is that this learned per-instance selection beats the best fixed single solver on the held-out 300-instance test set in normalized cost under both a 10-second and a 60-second budget, with a large and statistically significant gain at 10 seconds. The paper systematically compares cost-based, rank-based, and hybrid training losses and finds that the best objective depends on the budget: mean absolute error combined with a ranking loss at 10 seconds, a robust cost loss combined with a listwise ranking loss at 60 seconds. The work matters because it removes hand-crafted instance features from TSP algorithm selection and formulates selection as cost prediction plus ranking rather than earlier binary classification. The authors label the statistical tests as post-selection and exploratory, and at 60 seconds most solvers tie on most instances, so the reported gain rests on a small informative subset.

What carries the argument

The load-bearing machinery is a residual gated graph convolutional network that consumes a TSP instance as a complete graph with city-coordinate node features and distance edge features, runs three gated message-passing layers, pools the node embeddings into one graph vector by attention, and decodes a five-dimensional vector of predicted normalized costs. The selector chooses the solver with the lowest predicted value. Training targets are per-instance standardized costs, so the model learns relative solver quality rather than absolute tour length; rank labels built from those costs feed pairwise, listwise, and metric-driven ranking losses. The study's main experimental variable is the loss

What would settle it

Hold out the 300 public instances as a true test set, choose the loss configuration and the best-fixed-solver baseline on a separate validation sample from the same instance generators, then rerun the one-sided signed-rank test on fresh test instances. If the 60-second comparison no longer favors the learned selector, or the 10-second margin shrinks below significance, the paper's claim of improvement at both budgets is falsified.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that a graph neural network trained on raw TSP instances can select among a five-solver portfolio well enough to improve on the single best solver in normalized cost. On the 300-instance held-out test set, the selected 10-second model reaches a mean normalized cost of -0.4912 versus -0.4481 for the best fixed solver, closing roughly 73% of the mean gap between that solver and an oracle, with a one-sided signed-rank result driven by 154 instances better, 15 worse, and 131 ties. Under the 60-second budget, the selected model improves mean normalized cost from -0.2048 to -0.2548, and the paper reports a nominally significant signed-rank comparison, but th

Load-bearing premise

The load-bearing premise is that the 300-instance test set can serve both to pick the best loss configuration (and the single-best-solver baseline) and to compute the significance tests without inflating the apparent improvement; at 60 seconds, where only 11 instances have any best-fixed-solver-to-oracle gap, this reuse is what makes the reported significant result possible.

Editorial extensions

If this is right

  • Under a 10-second budget, per-instance selection with the best trained model improves mean normalized cost over the best fixed solver by about 0.043, an effect the paper finds highly significant (154 better, 15 worse, 131 ties).
  • The same selector does not significantly beat the best fixed solver in average rank, so the demonstrated advantage at 10 seconds is specifically a cost advantage, not a ranking advantage.
  • At a 60-second budget, most test instances have all solvers reaching the same final cost; the reported gain comes almost entirely from 11 instances with a real best-fixed-solver-to-oracle gap, so the practical headroom for algorithm selection shrinks as runtime grows.
  • The learned selector assigns different solvers across instances rather than collapsing to the single best baseline, and its selection frequencies resemble the tie-aware oracle more closely than any one solver.
  • A low-dimensional view of the instance-solver cost matrix shows the five solvers occupy distinct performance regions, supporting the premise that instance-dependent structure worth exploiting exists in the benchmark.

Reading between the lines

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

  • A stricter evaluation protocol would likely erase the 60-second significance: if the loss configuration and the single-best-solver baseline were chosen on a separate validation set and only then applied to a fresh test set, 13 nonzero paired differences (9 better, 4 worse) would be a fragile basis for a significant result. The paper already concedes the tests are post-selection.
  • Because all labels are normalized within each instance, the trained model cannot express absolute tour quality; on a distribution of uniformly harder instances the selector would still return the least-bad solver, so a selector can look effective in relative terms while absolute costs degrade. This is an inference, not a paper claim.
  • The same raw-graph cost-prediction-plus-ranking recipe should transfer to other graph-structured optimization problems with a heterogeneous solver portfolio, provided per-instance cost traces under a fixed cutoff exist; the paper does not test this.
  • A testable extension the paper leaves open: replace fixed cutoffs with full anytime solver trajectories and a temporal model, which would allow mid-run solver switching rather than a single upfront choice.
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

4 major / 5 minor

Summary. The paper introduces GNNAS-TSP, a GNN-based algorithm selector for TSP that takes raw graph inputs, predicts per-solver normalized costs for a five-solver portfolio (Concorde, CLK, EAX, LKH, MAOS), and selects the solver with the lowest predicted cost. It compares fifteen cost-based, rank-based, and hybrid loss configurations under 10s and 60s budgets on 300 public test instances, selects the best configuration per budget from the test-set aggregate results, and then analyzes VBS-gap closure, solver-selection frequencies, low-dimensional performance structure, and paired Wilcoxon comparisons against the single best solver (SBS). The central claim is that the selected configurations improve over SBS in normalized solution cost at both budgets, with a substantial and statistically significant improvement at 10s.

Significance. If the empirical claim holds, the paper provides useful evidence that feature-free GNN-based algorithm selection can outperform a strong fixed TSP solver under fixed budgets, and it offers a systematic comparison of cost and ranking losses in this setting. Strengths include the public release of code and data, the breadth of the loss-function sweep, explicit tie-aware analysis (VBS-fractional and gap-closed ratios), and the unusual candor of Sections 4.2 and 4.5 in labeling the post-selection analyses as exploratory. However, the paper's confirmatory framing in the abstract and Section 4.8 is not supported by the reported protocol.

major comments (4)
  1. [§4.2, §4.5, Table 4, Abstract] The same 300-instance test set is used both to select the best loss configuration per budget (Table 2) and to compute the paired Wilcoxon tests (Table 4). The reported p-values are therefore the best outcomes among fifteen configurations, not valid confirmatory inferences. The authors acknowledge this in §4.5, but the abstract and §4.8 present p=2.58e-20 (10s) and p=0.0374/0.0335 (60s) as statistical significance. The 60s result is especially fragile: it rests on only 13 non-tied pairs (9 better, 4 worse). A truly held-out validation split or nested model selection is required for confirmatory significance claims; otherwise the abstract and conclusions must be rephrased to describe only descriptive, exploratory improvements on the test sample.
  2. [§4.2] Each of the fifteen configurations is trained once, with no multiple seeds. The selection among close aggregate means (e.g., 10s MAE+LambdaRank at -0.4912 vs. MSE at -0.4803 vs. MAE+ListNet at -0.4786) could easily reflect optimization noise rather than a real loss-function advantage. This compounds the post-selection issue. The authors should report results over at least several independent training runs, or provide variance estimates, before claiming that one loss configuration is the 'best overall' or that the improvement magnitude is stable.
  3. [§3.1, §4.3] The primary evaluation metric is a per-instance z-score of solver costs (normalized relative to the mean and standard deviation of the five solvers on that instance). This is a relative standing measure, not an absolute tour-quality measure. The headline margins (0.0431 and 0.0500) are in standardized units and have no direct interpretation in terms of actual TSP solution cost. Since the stated goal is minimizing final solution cost, the paper should either report absolute tour costs or ratios to optimal/best-known values as a complement, or explicitly restrict the claim to 'relative normalized cost.'
  4. [§4.6, Table 5] The 60s improvement over SBS is driven almost entirely by 11 instances in the SBS>VBS subset. On the 289 SBS=VBS instances, Table 5 shows the selector's mean cost (-0.254) is actually worse than SBS (-0.272), even though most instances are exact ties. This context is described in §4.6, but the abstract and Section 4.8 state the 60s result as 'statistically significant' without this caveat. The conclusion should foreground that the 60s benefit is concentrated in a small informative subset, and that on the tied majority the selector does not improve over SBS.
minor comments (5)
  1. [Abstract / §4.8] The language 'statistically significant' is used without the exploratory caveat that is properly stated in §4.2 and §4.5. Please make the abstract and Section 4.8 consistent with the post-selection status of the Wilcoxon tests.
  2. [§3.3] The sentence 'We define the scores as the negative predicted costs because lower costs indicate better performance' appears twice in the RankNet paragraph. Please remove the duplicate.
  3. [Figure 2] The label 'Metasolver' appears to be a typo for 'Meta-solver.' Also, the arrow labeled 'Improving solution trajectories' is vague; clarify what is plotted or recorded.
  4. [Table 3] The 'Prvector' column is defined only in the table caption. Define it in the main text as well, since it is used to interpret the gap-closed behavior.
  5. [§3.1] The tie-breaker rule for training ranks (runtime used only for equal costs) is later said to make RankNet learn 'secondary runtime preference.' This is fine, but it should be flagged more explicitly as a design choice with possible consequences for the rank-based losses.

Circularity Check

1 steps flagged · score 2.0 of 10

No definitional circularity; the GNN-AS result is externally grounded, but the reported Wilcoxon significance is post-selection because the same 300 test instances were used to choose the loss configuration.

  1. fitted input called prediction [Section 4.2 and Section 4.5; echoed in Abstract and Section 4.8]
    "We use the aggregate test-set results in Table 2 to identify one representative high-performing loss configuration for each time budget. ... The paired Wilcoxon analyses below use the same test instances and therefore constitute post-selection exploratory analyses rather than independent confirmatory tests."

    The same 300-instance test set is used twice in the inference chain: first as the selection criterion for the best loss configuration (MAE+LambdaRank at 10s, Huber+ListNet at 60s), then as the data for the one-sided Wilcoxon tests in Table 4. The reported p-values are therefore conditional on having chosen the configuration that looked best on those exact instances—they are post-selection statistics, not pre-specified confirmatory tests. The abstract's 'statistically significant cost improvement' and Section 4.8's p=2.58e-20 / p=0.0374 / p=0.0335 thus describe the same data used for selection rather than an independent prediction. The 60s result is especially fragile because only 13 non-tied pairs (9 better, 4 worse) drive the p-value. This does not erase the descriptive improvement on the

full rationale

The paper's core derivation is not circular: GNNAS-TSP is trained on 700 instances with solver-cost labels from external TSP benchmarks and evaluated on 300 disjoint test instances; the cost/rank losses are standard and the selection rule is a direct argmin over predicted costs. There is no self-citation chain, uniqueness theorem, or ansatz smuggled via citation that forces the conclusion. The only notable issue is the post-selection use of the held-out test set: Section 4.2 explicitly says Table 2 is used to identify the best configuration, and Section 4.5 explicitly labels the Wilcoxon tests as post-selection and exploratory, yet the Abstract and Section 4.8 present those same p-values as statistically significant support. This is a methodological selection-bias concern rather than a definitional equivalence, so it warrants a mild score of 2 rather than a higher circularity score. The descriptive claim that the selected configurations improve on SBS in normalized cost on this test sample remains externally grounded and reproducible from the reported data.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central empirical result rests on standard GNN architectures and a public dataset; the main free choices are the loss configuration, hyperparameters, and the evaluation protocol. No invented physical entities.

free parameters (4)
  • Best loss configuration per budget = MAE+LambdaRank (10s), Huber+ListNet (60s)
    Selected using aggregate cost/rank on the same 300-instance test set (Table 2, Section 4.2); this categorical free choice directly determines the reported improvement.
  • Learning rate decay and max epochs = decay 1.2 every 5 epochs; 30 epochs
    Chosen by a 'small one-factor sensitivity check' (Section 4.2); the validation set for this check is not specified, so it may have been the test set.
  • Architecture hyperparameters = h=128, 3 conv layers, 2 MLP layers
    Fixed following related studies (Kool et al., Song et al., Joshi et al.) rather than tuned; still a modeling choice affecting capacity.
  • Hybrid loss weight = 0.5 cost + 0.5 rank
    Fixed arbitrarily for hybrid losses; no ablation of this weight is reported.
assumptions (5)
  • domain assumption Solver performance traces in the Huerta et al. dataset are accurate at 10s/60s cutoffs.
    The entire label construction depends on these traces (Section 3.1).
  • domain assumption A complete-graph ResGatedGCN with 3 layers can learn instance-dependent solver performance from node coordinates.
    The method's central hypothesis (Section 3.2).
  • standard math Per-instance standardization preserves the argmin selection and enables cross-instance training.
    Monotone transform per instance, so the lowest standardized cost corresponds to the lowest true cost (Section 3.1).
  • domain assumption The training distribution (RUE/NETGEN/NETGENM/TSPGEN) transfers to the test distribution (TSPLIB/TNM/NATIONAL/VLSI).
    Cross-distribution generalization is assumed without analysis (Sections 3.1, 4.2).
  • domain assumption The 300 test instances are representative enough to support the significance claims.
    Used for model selection and inference (Section 4.5).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Graph Neural Network-based Algorithm Selection for the Traveling Salesman Problem: A Systematic Study of Cost and Rank Losses under Distinct Budget Regimes." pith.science (2026). https://pith.science/paper/4LHBOAQH

@misc{pith2026260718632,
  author       = {Pith},
  title        = {Pith review of: Graph Neural Network-based Algorithm Selection for the Traveling Salesman Problem: A Systematic Study of Cost and Rank Losses under Distinct Budget Regimes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4LHBOAQH}},
  note         = {Machine review of arXiv:2607.18632}
}
read the original abstract

Automated Algorithm Selection (AS) aims to improve problem-solving performance by selecting, for each problem instance, the most suitable algorithm from a predefined portfolio. This is particularly relevant to the Traveling Salesman Problem (TSP), where solver performance is strongly instance-dependent. We introduce GNNAS-TSP, a Graph Neural Network (GNN)-based AS framework that learns TSP instance representations directly from raw graph data, avoiding manual feature engineering. GNNAS-TSP formulates AS as a joint cost-prediction and ranking task. We evaluate cost-based (mean squared error (MSE), mean absolute error (MAE), and Huber), rank-based (RankNet, ListNet, and LambdaRank), and hybrid learning objectives for a portfolio comprising Chained Lin-Kernighan, Edge Assembly Crossover, Lin-Kernighan-Helsgaun, Multiagent Optimization System, and Concorde. Experiments use fixed computational budgets of 10 and 60 seconds. On the held-out test set, the selected configurations improve on the Single Best Solver (SBS) in normalized solution cost at both budgets. For the 10s budget, AS achieves substantial and statistically significant cost improvement over SBS. Overall, the results suggest that GNNAS-TSP is a useful meta-solving strategy when exploitable variation exists across solver performance.

Figures

Figures reproduced from arXiv: 2607.18632 by the authors.

Figure 1
Figure 1. A standard Algorithm Selection (AS) pipeline: represent the instance, predict relative solver perfor [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The high-level setup of the GNNAS-TSP framework. The [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Model architecture. The input TSP graph G = (X, A, E) is encoded via ResGatedGCN to produce node embeddings H(L) . These are aggregated into a graph-level representation hG via global attention pooling, which is then decoded by an MLP to predict the normalized costs or ranks yˆ ∈ R m for the TSP solver portfolio P = {A1, . . . , Am}. where edges are possible between any pair of cities in a route. However, since opti… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Solver selection frequencies as percentages over 300 test instances for the 10-second (left) and 60-second [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: SVD visualization of the 10s normalized-cost matrix. Left: instances colored by hard VBS solver. [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: SVD visualization of the 60s normalized-cost matrix. Left: instances colored by hard VBS solver. [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

77 extracted references · 12 linked inside Pith

  1. [1]

    Encyclopedia of operations research and management science , volume=

    Traveling salesman problem , author=. Encyclopedia of operations research and management science , volume=. 2013 , publisher=

    2013

  2. [2]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Learning to solve np-complete problems: A graph neural network for decision tsp , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  3. [3]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Dualopt: A dual divide-and-optimize algorithm for the large-scale traveling salesman problem , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  4. [4]

    arXiv preprint arXiv:2501.14285 , year=

    Cascaded large-scale tsp solving with unified neural guidance: Bridging local and population-based search , author=. arXiv preprint arXiv:2501.14285 , year=

    arXiv

  5. [5]

    Pattern Recognition , pages=

    GELD: A unified neural model for efficiently solving traveling salesman problems across different scales , author=. Pattern Recognition , pages=. 2025 , publisher=

    2025

  6. [6]

    arXiv preprint arXiv:2509.23465 , year=

    ViTSP: A Vision Language Models Guided Framework for Large-Scale Traveling Salesman Problems , author=. arXiv preprint arXiv:2509.23465 , year=

    arXiv

  7. [7]

    Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

    Diversity optimization for travelling salesman problem via deep reinforcement learning , author=. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1 , pages=

  8. [8]

    Nature Communications , year=

    Probabilistic greedy algorithm solver using magnetic tunneling junctions for traveling salesman problem , author=. Nature Communications , year=

Show all 77 references
  1. [9]

    arXiv preprint arXiv:2504.15147 , year=

    The iterative chainlet partitioning algorithm for the traveling salesman problem with drone and neural acceleration , author=. arXiv preprint arXiv:2504.15147 , year=

  2. [10]

    Neurocomputing , pages=

    Heterogeneous graph neural networks for scalable asymmetric traveling salesman problem optimization , author=. Neurocomputing , pages=. 2026 , publisher=

    2026

  3. [11]

    2025 RIVF International Conference on Computing and Communication Technologies (RIVF) , pages=

    NeuFACO: Neural Focused Ant Colony Optimization for Traveling Salesman Problem , author=. 2025 RIVF International Conference on Computing and Communication Technologies (RIVF) , pages=. 2025 , organization=

    2025

  4. [12]

    Journal of Statistical Software , volume=

    TSP—infrastructure for the traveling salesperson problem , author=. Journal of Statistical Software , volume=

  5. [13]

    arXiv preprint arXiv:1806.09285 , year=

    A new benchmark set for Traveling salesman problem and Hamiltonian cycle problem , author=. arXiv preprint arXiv:1806.09285 , year=

    arXiv

  6. [14]

    URL: http://comopt

    \ TSPLIB \ : a library of sample instances for the TSP (and related problems) from various sources and of various types , author=. URL: http://comopt. ifi. uniheidelberg. de/software/TSPLIB95 , year=

  7. [15]

    ORSA journal on computing , volume=

    TSPLIB—A traveling salesman problem library , author=. ORSA journal on computing , volume=. 1991 , publisher=

    1991

  8. [16]

    2007 International Conference on Machine Learning and Cybernetics , volume=

    An improved genetic algorithm for tsp , author=. 2007 International Conference on Machine Learning and Cybernetics , volume=. 2007 , organization=

    2007

  9. [17]

    Proceedings of the Genetic and Evolutionary Computation Conference , pages=

    Improving an exact solver for the traveling salesman problem using partition crossover , author=. Proceedings of the Genetic and Evolutionary Computation Conference , pages=

  10. [18]

    Asian conference on machine learning , pages=

    Learning 2-opt heuristics for the traveling salesman problem via deep reinforcement learning , author=. Asian conference on machine learning , pages=. 2020 , organization=

    2020

  11. [19]

    arXiv preprint arXiv:1611.09940 , year=

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

    arXiv

  12. [20]

    arXiv preprint arXiv:1803.08475 , year=

    Attention, learn to solve routing problems! , author=. arXiv preprint arXiv:1803.08475 , year=

    arXiv

  13. [21]

    2006 , publisher =

    The Traveling Salesman Problem: A Computational Study , author =. 2006 , publisher =

    2006

  14. [22]

    INFORMS Journal on Computing , volume =

    Chained Lin--Kernighan for Large Traveling Salesman Problems , author =. INFORMS Journal on Computing , volume =

  15. [23]

    Advances in computers , volume=

    The algorithm selection problem , author=. Advances in computers , volume=. 1976 , publisher=

    1976

  16. [24]

    AI Magazine , volume =

    Kotthoff, Lars , title =. AI Magazine , volume =. 2014 , doi =

    2014

  17. [25]

    Journal of artificial intelligence research , volume=

    SATzilla: portfolio-based algorithm selection for SAT , author=. Journal of artificial intelligence research , volume=

  18. [26]

    IEEE transactions on neural networks , volume=

    The graph neural network model , author=. IEEE transactions on neural networks , volume=. 2008 , publisher=

    2008

  19. [27]

    IEEE transactions on neural networks and learning systems , volume=

    A comprehensive survey on graph neural networks , author=. IEEE transactions on neural networks and learning systems , volume=. 2020 , publisher=

    2020

  20. [28]

    Artificial Intelligence , volume=

    Alors: An algorithm recommender system , author=. Artificial Intelligence , volume=. 2017 , publisher=

    2017

  21. [29]

    Ieee Access , volume=

    A graph neural network assisted monte carlo tree search approach to traveling salesman problem , author=. Ieee Access , volume=. 2020 , publisher=

    2020

  22. [30]

    Distill , volume=

    A gentle introduction to graph neural networks , author=. Distill , volume=

  23. [31]

    arXiv preprint , year=

    Combinatorial Optimization with Automated Graph Neural Network , author=. arXiv preprint , year=

  24. [32]

    Journal of Heuristics , volume=

    A case study of algorithm selection for the traveling thief problem , author=. Journal of Heuristics , volume=

  25. [33]

    arXiv preprint , year=

    GNNAS-Dock: Budget Aware Algorithm Selection with Graph Neural Networks for Molecular Docking , author=. arXiv preprint , year=

  26. [34]

    Advances in computers , volume=

    The algorithm selection problem , author=. Advances in computers , volume=

  27. [35]

    2021 , url=

    Michael Atkin , title=. 2021 , url=

    2021

  28. [36]

    arXiv preprint arXiv:2110.05291 , year=

    Graph neural network guided local search for the traveling salesperson problem , author=. arXiv preprint arXiv:2110.05291 , year=

    arXiv

  29. [37]

    Advances in Neural Information Processing Systems , volume=

    Unsupervised learning for solving the travelling salesman problem , author=. Advances in Neural Information Processing Systems , volume=

  30. [38]

    International Conference on Algorithmic Aspects in Information and Management , pages=

    Quantum Annealing and GNN for Solving TSP with QUBO , author=. International Conference on Algorithmic Aspects in Information and Management , pages=. 2024 , organization=

    2024

  31. [39]

    Travelling Salesman Problem Goes Sparse With Graph Neural Networks , author=

  32. [40]

    Advances in neural information processing systems , volume=

    How does batch normalization help optimization? , author=. Advances in neural information processing systems , volume=

  33. [41]

    AI Open , volume=

    Graph neural networks: A review of methods and applications , author=. AI Open , volume=

  34. [42]

    arXiv preprint , year=

    A Unified Framework for Combinatorial Optimization Based on Graph Neural Networks , author=. arXiv preprint , year=

  35. [43]

    arXiv preprint arXiv:2103.14807 , year=

    An Introduction to Robust Graph Convolutional Networks , author=. arXiv preprint arXiv:2103.14807 , year=

    arXiv

  36. [44]

    arXiv preprint arXiv:2404.13079 , year=

    Relational Graph Convolutional Networks for Sentiment Analysis , author=. arXiv preprint arXiv:2404.13079 , year=

    arXiv

  37. [45]

    Future Internet , volume=

    Graph attention networks: a comprehensive review of methods and applications , author=. Future Internet , volume=. 2024 , publisher=

    2024

  38. [46]

    arXiv preprint , year=

    An Efficient Graph Convolutional Network Technique for the Travelling Salesman Problem , author=. arXiv preprint , year=

  39. [47]

    Expert Systems With Applications , volume=

    Improving the state-of-the-art in the Traveling Salesman Problem: An Anytime Automatic Algorithm Selection , author=. Expert Systems With Applications , volume=

  40. [48]

    European Journal of Operational Research , volume=

    An effective implementation of the Lin-Kernighan traveling salesman heuristic , author=. European Journal of Operational Research , volume=

  41. [49]

    IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) , volume=

    Multiagent optimization system for solving the traveling salesman problem (TSP) , author=. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) , volume=. 2008 , publisher=

    2008

  42. [50]

    INFORMS Journal on Computing , volume=

    A Powerful Genetic Algorithm Using Edge Assembly Crossover for the Traveling Salesman Problem , author=. INFORMS Journal on Computing , volume=

  43. [51]

    arXiv preprint arXiv:2302.04035 , year=

    Revisit the algorithm selection problem for TSP with spatial information enhanced graph neural networks , author=. arXiv preprint arXiv:2302.04035 , year=

    arXiv

  44. [52]

    17th International Conference on Agents and Artificial Intelligence, ICAART 2025 , pages=

    Revisit the Algorithm Selection Problem for TSP with Spatial Information Enhanced Graph Neural Networks , author=. 17th International Conference on Agents and Artificial Intelligence, ICAART 2025 , pages=. 2025 , organization=

    2025

  45. [53]

    International Conference on Learning and Intelligent Optimization , pages=

    Geometrically Invariant and Equivariant Graph Neural Networks for TSP Algorithm Selection and Hardness Prediction , author=. International Conference on Learning and Intelligent Optimization , pages=. 2025 , organization=

    2025

  46. [54]

    Applied Soft Computing , pages=

    Meta-learning-based adaptive operator selection for traveling salesman problem , author=. Applied Soft Computing , pages=. 2025 , publisher=

    2025

  47. [55]

    arXiv preprint arXiv:2509.26377 , year=

    MC-GNNAS-Dock: Multi-criteria GNN-based Algorithm Selection for Molecular Docking , author=. arXiv preprint arXiv:2509.26377 , year=

  48. [56]

    Biometrics bulletin , volume=

    Individual comparisons by ranking methods , author=. Biometrics bulletin , volume=. 1945 , publisher=

    1945

  49. [57]

    com: A tool for life science researchers for multiple hypothesis testing correction , author=

    MultipleTesting. com: A tool for life science researchers for multiple hypothesis testing correction , author=. PloS one , volume=. 2021 , publisher=

    2021

  50. [58]

    Scandinavian journal of statistics , pages=

    A simple sequentially rejective multiple test procedure , author=. Scandinavian journal of statistics , pages=. 1979 , publisher=

    1979

  51. [59]

    Proceedings of the aaai conference on artificial intelligence , volume=

    Deep learning for algorithm portfolios , author=. Proceedings of the aaai conference on artificial intelligence , volume=

  52. [60]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Evolvegcn: Evolving graph convolutional networks for dynamic graphs , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  53. [61]

    arXiv preprint arXiv:2006.10637 , year=

    Temporal graph networks for deep learning on dynamic graphs , author=. arXiv preprint arXiv:2006.10637 , year=

    2006 arXiv

  54. [62]

    Evolutionary computation , volume=

    Automated algorithm selection: Survey and perspectives , author=. Evolutionary computation , volume=. 2019 , publisher=

    2019

  55. [63]

    Journal of Heuristics , volume=

    A case study of algorithm selection for the traveling thief problem , author=. Journal of Heuristics , volume=. 2018 , publisher=

    2018

  56. [64]

    IEEE Transactions on Evolutionary Computation , volume=

    No free lunch theorems for optimization , author=. IEEE Transactions on Evolutionary Computation , volume=. 1997 , publisher=

    1997

  57. [65]

    Artificial Intelligence Review , volume=

    Solving the traveling salesman problem with machine learning: a review of recent advances and challenges , author=. Artificial Intelligence Review , volume=. 2025 , publisher=

    2025

  58. [66]

    Artificial Intelligence , volume=

    Bischl, Bernd and Kerschke, Pascal and Kotthoff, Lars and Lindauer, Marius and Malitsky, Yuri and Fr. Artificial Intelligence , volume=. 2017 , publisher=

    2017

  59. [67]

    Proceedings of the 9th Learning and Intelligent OptimizatioN Conference (LION) , volume =

    Improving the State of the Art in Inexact TSP Solving using Per-Instance Algorithm Selection , author=. Proceedings of the 9th Learning and Intelligent OptimizatioN Conference (LION) , volume =. 2015 , organization=

    2015

  60. [68]

    IEEE Congress on Evolutionary Computation (CEC) , year=

    Outperforming the best with minimal effort: Algorithm selection for constrained multi-objective optimization , author=. IEEE Congress on Evolutionary Computation (CEC) , year=

  61. [69]

    2005 , isbn =

    Burges, Chris and Shaked, Tal and Renshaw, Erin and Lazier, Ari and Deeds, Matt and Hamilton, Nicole and Hullender, Greg , title =. 2005 , isbn =. doi:10.1145/1102351.1102363 , booktitle =

    2005

  62. [70]

    2007 , isbn =

    Cao, Zhe and Qin, Tao and Liu, Tie-Yan and Tsai, Ming-Feng and Li, Hang , title =. 2007 , isbn =. doi:10.1145/1273496.1273513 , booktitle =

    2007

  63. [71]

    Learning to Rank with Nonsmooth Cost Functions , url =

    Burges, Christopher and Ragno, Robert and Le, Quoc , booktitle =. Learning to Rank with Nonsmooth Cost Functions , url =

  64. [72]

    arXiv preprint arXiv:2512.02328 , year=

    Molecular Embedding-Based Algorithm Selection in Protein-Ligand Docking , author=. arXiv preprint arXiv:2512.02328 , year=

    arXiv

  65. [73]

    Journal of Cheminformatics , year=

    Molecular embedding-based algorithm selection in protein-ligand docking , author=. Journal of Cheminformatics , year=

  66. [74]

    European Conference on Evolutionary Computation in Combinatorial Optimization (Part of EvoStar) , pages=

    Algorithm selection for the team orienteering problem , author=. European Conference on Evolutionary Computation in Combinatorial Optimization (Part of EvoStar) , pages=. 2022 , organization=

    2022

  67. [75]

    Theoretical Computer Science , volume=

    A study on the effects of normalized TSP features for automated algorithm selection , author=. Theoretical Computer Science , volume=. 2023 , publisher=

    2023

  68. [76]

    arXiv preprint arXiv:2009.03509 , year=

    Masked label prediction: Unified message passing model for semi-supervised classification , author=. arXiv preprint arXiv:2009.03509 , year=

    2009 arXiv

  69. [77]

    arXiv preprint arXiv:2605.20797 , year=

    Beyond Numerical Features: CNN-Driven Algorithm Selection via Contour Plots for Continuous Black-Box Optimization , author=. arXiv preprint arXiv:2605.20797 , year=

    arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.