Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

An End-to-End Learning Approach for Solving Capacitated Location-Routing Problems

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper claims that capacitated location-routing problems can be solved end-to-end by one attention-based policy, jointly deciding where to open depots and how to route vehicles, and that it outperforms traditional heuristics and prior DR

desk verdict A useful transferable MDP/masking framework for CLRP with strong empirical results, but the feasibility guarantee is broken by a simple dead-end counterexample. read the letter →

arxiv 2511.02525 v2 pith:WK5WD3SN submitted 2025-11-04 cs.LG cs.AI

Changhao Miao , Yuntian Zhang , Tongyu Wu , Fang Deng , Chen Chen This is my paper · ORCID
classification cs.LGcs.AI
keywords location-routingproblemdeepreinforcementlearningend-to-endattentionmechanismcapacitatedvehicleroutingMDPformulationdynamicmaskingcombinatorialoptimization
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 sets out to show that the capacitated location-routing problem (CLRP) and its open-route variant (OCLRP) need not be decomposed into separate location and routing subproblems. It proposes DRLHQ, a deep reinforcement learning approach that constructs a full solution node by node in a single pass, alternating between choosing a depot to open and choosing customers to visit. The authors argue that this end-to-end treatment respects the coupling between facility location and routing, and they introduce a heterogeneous querying attention mechanism and a dynamic masking scheme to make it work. If the claims hold, DRL-based planners could handle combined location-routing decisions directly, with better solution quality and faster inference than classic exact and heuristic methods. The core evidence is a set of experiments on synthetic and benchmark instances where DRLHQ posts gaps of 7.59% (CLRP) and 6.08% (OCLRP) to best-known solutions, roughly half the gap of the strongest DRL baselines.

What carries the argument

The load-bearing object is the heterogeneous querying attention mechanism, driven by an MDP with an indicator state I_t that toggles between location and routing decisions. At location steps, a GRU converts the previous depot's embedding into a location query; at routing steps, the context embedding of the last visited node becomes the routing query. A dynamic masking mechanism with five rules enforces feasibility: it masks visited customers, non-departure depots during a subtour, customers exceeding vehicle/depot residual capacity, and depots whose remaining capacity cannot cover the smallest remaining demand. These pieces allow a single encoder-decoder policy to construct entire CLRP solut

What would settle it

Construct a small instance with two depots and three customers where one customer's demand exceeds the residual capacity of every depot after the other two customers are served greedily, while that customer's demand is above the minimum demand used in Masking Rule 5. Run the trained DRLHQ policy with greedy decoding and check whether it terminates with a feasible solution; if it dead-ends or must violate the depot capacity constraint, the feasibility guarantee in the paper is incomplete.

Watch

Extended reading notes

Core claim

The central claim is that a single policy network, trained end-to-end with REINFORCE and a shared baseline, can produce feasible, high-quality CLRP solutions by interleaving two kinds of decisions within one sequential construction process. The key is an MDP reformulation in which an indicator state switches between 'location' actions (pick the starting depot for the next subtour) and 'routing' actions (pick the next customer or return to depot), synchronized by dynamic masking rules that enforce vehicle and depot capacity constraints. On top of this, a heterogeneous querying attention mechanism builds a GRU-based location query for depot choices and a context-based routing query for custome

Load-bearing premise

The claim that solutions are always feasible rests on Masking Rule 5, which assumes a depot can be ruled out just by comparing its remaining capacity to the smallest unserved demand; no look-ahead guards against stranding a large-demand customer when all remaining depot capacities have been partly consumed.

Editorial extensions

If this is right

  • Adapting the dynamic masking mechanism alone substantially improves prior DRL methods: POMO's gap on CLRP100 drops from 26.91% to 1.79% when retrofitted with the masking rules.
  • The end-to-end policy generalizes across scale: a model trained on 100 customers continues to outperform heuristics and DRL baselines on 125-200 customer instances without retraining.
  • Instance augmentation by rotation and simulation-based beam search are drop-in decoding enhancements that further reduce gap, up to 2.31% on CLRP50 with SBS.
  • Benchmark results imply that a purely learning-based approach can come within 6-8% of best-known solutions on published instances, while computing solutions in under a second for 100-customer problems.

Reading between the lines

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

  • The MDP indicator-state formulation is a template: the same alternating-decision loop could be applied to other combined problems such as inventory-location-routing or multi-echelon distribution, without redesigning the policy head, as long as appropriate masking rules are defined.
  • The greedy masking rule checks only the minimum remaining demand, not the largest; a look-ahead feasibility test (masking a depot if its residual capacity cannot cover the maximum unserved demand) would make the construction feasible by construction and might remove rare dead-ends that the current method must implicitly avoid during training.
  • The heterogeneous query design suggests a broader principle: when one policy must produce structurally different kinds of decisions, giving each decision type its own query representation is more effective than a single context vector. This could transfer to other construction heuristics that mix assignment, sequencing, and resource-allocation decisions.
  • Because the policy is trained purely on uniformly sampled synthetic instances, the 6-8% gap on real-world benchmarks may be shrinkable with light fine-tuning on the benchmark distribution or with a learned repair operator, which the paper does not explore.
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 proposes DRLHQ, a deep reinforcement learning method for the capacitated location-routing problem (CLRP) and its open-route variant (OCLRP). The method reformulates the problem as an MDP with alternating location and routing decisions, uses an attention-based encoder-decoder with a heterogeneous query mechanism (a GRU-based location query and a routing query), and applies a dynamic masking mechanism to enforce constraints. Training uses REINFORCE with a POMO-style shared baseline; inference uses a rotation-based instance augmentation and a simulation-based beam search. The paper claims to be the first end-to-end DRL approach for CLRPs and reports superior solution quality and better generalization than Gurobi, ILS, SAH, TBSA, AM-W, POMO, and MTA on synthetic and public benchmark instances.

Significance. If the claims are substantiated, the paper would make a useful contribution: an end-to-end constructive policy for CLRP/OCLRP that avoids the two-phase decomposition of earlier DRL work, a reusable MDP formulation, and strong empirical results on benchmark instances. The adaptation of the masking mechanism to POMO/MTA and the reported ablation study are also valuable. However, the central claims currently rest on a feasibility guarantee that is not established, and the OCLRP cost model appears to undercount route costs. The novelty claim also needs to be delimited against the cited related work. The paper has no machine-checked proofs or code release; its strengths are the breadth of experiments and the public-benchmark evaluation, but those need to be expressed with proper statistical and feasibility checks.

major comments (4)
  1. [§IV.D, Masking Rule 5 and Eqs. (16)/(18)] The claim that the dynamic masking mechanism 'ensures feasibility' is not supported. Rule 5 only masks a depot when its remaining capacity is below the minimum remaining customer demand; it never checks whether every unserved customer can still be assigned to some depot with enough remaining capacity. Consider two depots of capacity 100, customers with demands {100,1,1}, and vehicle capacity 100. The sequence 'open depot 1, serve a demand-1 customer, return; open depot 2, serve the other demand-1 customer, return' is allowed by Rules 1–5. The remaining demand is 100 while both depots have 99 capacity left, so Rule 5 masks all depots and the construction dead-ends on a feasible instance. If such states are reachable, the reward in Eq. (30) and the objective values in Tables I–V are undefined for those trajectories. Please add a global feasibility check/repair, or report exact dead-end/inf
  2. [§III, Eq. (15) and §IV-A] The OCLRP cost model appears wrong. The paper sets c_{ji}=0 for all i∈I,j∈J. Because the graph is undirected and c_{ij}=c_{ji}, this makes both the customer-to-depot return edge and the depot-to-first-customer departure edge cost zero. An open route still incurs the cost of traveling from the depot to its first customer; only the final return edge should be removed. As written, Eq. (1) and the transition cost in Eq. (17) undercount every OCLRP route by one depot–customer arc. The OCLRP experiments in Tables II and V therefore solve a looser objective. Please correct the formulation (e.g., directed zero-cost return arcs) and rerun the OCLRP comparisons.
  3. [Abstract/§II-C] The 'first end-to-end' claim is too broad. The related-work section itself cites [37] as embedding neural networks 'to handle both location and routing tasks simultaneously' and [38] as a two-stage attention model for the Two-Echelon Location-Routing Problem. No definition of 'end-to-end' is given that would exclude these methods. Please delimit the novelty claim precisely (e.g., single integrated constructive policy for CLRP/OCLRP with no separate optimization loop) and explicitly explain why [37] and [38] do not satisfy that definition.
  4. [§V-A/B, Tables I–III] The empirical support for the 'superior solution quality' claim is incomplete. All numbers are point estimates over 1,000 instances with no standard deviations, confidence intervals, or paired tests. Several claimed wins are sub-1% (e.g., Table I, CLRP10: DRLHQ-Aug-8 8.6499 vs MTA*-Aug-8 8.6527), so the separation is not established. In Table V, one reported POMO* solution (100-10-1, 263202) is 4.93% below the stated BKS, which indicates the benchmark BKS values are inconsistent. In addition, the SBS results in Table III do not report the beam width β or expansion factor γ used, and no code is provided; the experiments are not reproducible from the manuscript.
minor comments (5)
  1. [§IV-A] The text says 'as indicated in Eq. 14' when referring to the OCLRP zero-distance rule; the correct reference is Eq. (15).
  2. [§IV-C] Typo: 'contacting' should be 'concatenating' in the description of the context embedding.
  3. [Eq. (28)] The attention equation is structurally unclear. As written, it places the value V inside the tanh and does not explicitly show a softmax over attention scores before weighting. Please align the notation with standard multi-head attention or [34].
  4. [Table I] The Gurobi(3600s) row appears misaligned: it lacks the CLRP10 columns and seems to have shifted values into the CLRP20/50/100 columns.
  5. [§IV-F] The statement that 'flipping or rotating an instance does not change its optimal solution' should be phrased as isometric equivalence: the optimal route is rotated/flipped and the cost is invariant, not that the same optimal solution remains unchanged.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the central pipeline is empirical, parameter-free with respect to the test benchmarks, and the only self-citation is a background pointer.

full rationale

DRLHQ defines an MDP, an encoder-decoder policy, and dynamic masking rules, then trains with REINFORCE on synthetic instances. No equation in Section IV is fitted to the BKS values in Tables IV–V, and no benchmark result is inserted as a training target; the rotation augmentation in Eq. (32) is an invariance transformation, not a data-dependent fit. The comparisons against Gurobi, ILS, SAH, TBSA, AM-W, POMO, and MTA are external, and the benchmark gaps are measured against external best-known solutions from Prins et al. [43], so the reported improvements are not true by construction. The self-citation [13] in I.A and II.B is used only as a general pointer for DRL on facility-location problems; it is not the basis of the MDP reformulation, masking rules, or any uniqueness or optimality claim, so it is not load-bearing. The skeptic's concern about Masking Rule 5 (Section IV.D) is a soundness/feasibility gap — a local capacity check may in principle dead-end on feasible instances — but that is a correctness risk, not a circular reduction of the method's predictions to its inputs. Accordingly no circular step is identified; the score is 1 only because of the minor non-load-bearing self-citation.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The report's central claim is empirical, so the ledger contains no fitted physical constants. The hand-chosen items are decoding/search hyperparameters and four assumptions about instance geometry, feasibility, and OCLRP cost structure. The most consequential is the local-capacity masking rule, which is assumed to imply global feasibility without a proof or repair mechanism.

free parameters (3)
  • Simulation-based beam search width beta and expansion factor gamma = not stated
    Algorithm 2 defines SBS with beta and gamma; Table III results depend on them but their values are never reported.
  • Number of instance augmentations g = 8 (also 4 tested)
    Inference uses g rotations; larger g improves objective in Tables I-II at extra compute. This is a search-effort knob, not a fitted physical constant.
  • Architecture/training hyperparameters (L=6, H=16, d_h=256, FF=512, tanh clip=10, lr schedule) = as in Section V.A
    Borrowed from [16]/[34]; chosen by hand and influencing all reported numbers, but not fitted to the benchmark.
assumptions (5)
  • standard math Edge costs are symmetric and satisfy the triangle inequality.
    Stated in Section III; supports cost computation c_ij in the objective and route feasibility.
  • domain assumption Every tested CLRP/OCLRP instance is globally feasible (total depot capacity can cover total demand).
    Not stated; dynamic masking only enforces local capacity checks (Masking Rules 3 and 5), so construction can dead-end if earlier depot selections consume capacity needed by a large remaining customer.
  • ad hoc to paper For OCLRP, setting customer-to-depot edge cost to zero implements open routes.
    Eq. (15) c_ji=0 for i in I, j in J is inconsistent with the undirected graph definition c_ij=c_ji>0 and would zero both depot-customer directions if read literally.
  • ad hoc to paper Rotational augmentation preserves the optimal solution and model behavior for arbitrary angles.
    Algorithm 1 and Eq. (32) claim arbitrary rotation invariance; for coordinates uniformly sampled in the unit square, rotations other than multiples of 90 degrees move points outside the training distribution, so the general claim is not justified.
  • domain assumption A CLRP solution can be represented as a sequence of subtours, each using one vehicle and one depot.
    This is the standard CLRP decomposition used to build the MDP; it is not restrictive when the vehicle set is unlimited, but the paper does not prove the constructed MDP covers all optimal solutions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An End-to-End Learning Approach for Solving Capacitated Location-Routing Problems." pith.science (2026). https://pith.science/paper/WK5WD3SN

@misc{pith2026251102525,
  author       = {Pith},
  title        = {Pith review of: An End-to-End Learning Approach for Solving Capacitated Location-Routing Problems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WK5WD3SN}},
  note         = {Machine review of arXiv:2511.02525}
}
read the original abstract

The capacitated location-routing problems (CLRPs) are classical problems in combinatorial optimization, which require simultaneously making location and routing decisions. In CLRPs, the complex constraints and the intricate relationships between various decisions make the problem challenging to solve. With the emergence of deep reinforcement learning (DRL), it has been extensively applied to address the vehicle routing problem and its variants, while the research related to CLRPs still needs to be explored. In this paper, we propose the DRL with heterogeneous query (DRLHQ) to solve CLRP and open CLRP (OCLRP), respectively. We are the first to propose an end-to-end learning approach for CLRPs, following the encoder-decoder structure. In particular, we reformulate the CLRPs as a markov decision process tailored to various decisions, a general modeling framework that can be adapted to other DRL-based methods. To better handle the interdependency across location and routing decisions, we also introduce a novel heterogeneous querying attention mechanism designed to adapt dynamically to various decision-making stages. Experimental results on both synthetic and benchmark datasets demonstrate superior solution quality and better generalization performance of our proposed approach over representative traditional and DRL-based baselines in solving both CLRP and OCLRP.

Figures

Figures reproduced from arXiv: 2511.02525 by the authors.

Figure 1
Figure 1. An illustrative example of CLRP. The decision process of CLRP can be divided into three partitions: (a) Facility Location, (b) Customer Allocation, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overall pipeline of DRLHQ. We propose a heterogeneous querying attention mechanism that invokes distinct query vectors tailored to various [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Illustration of using our heterogeneous querying attention mechanism to determine an instance involving two depots and five customers. The upper [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The generalization results in larger scales for CLRP and OCLRP, [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Adversarial Training for Robust Coverage Network under Worst-case Facility Losses

    cs.LG 2026-05 unverdicted novelty 7.0 of 10

    A dual-agent adversarial DRL method solves the MCLIP bi-level problem with superior efficiency and competitive quality versus baselines on synthetic and real datasets.

  2. Towards Generalization-Oriented Models for Vehicle Routing Problems with Mixture-of-Experts

    cs.LG 2026-05 unverdicted novelty 6.0 of 10

    R2E-IG combines residual refined experts with instance-level gating and mixed-distribution training using dynamic weight adaptation to improve generalization of DRL solvers for vehicle routing problems.

Reference graph

Works this paper leans on

48 extracted references · 1 linked inside Pith · cited by 2 Pith papers

  1. [37]

    Neural embedded opti- mization for integrated location and routing problems,

    W. Kaleem, H. Ayala, and A. Subramanyam, “Neural embedded opti- mization for integrated location and routing problems,” inIISE Annual Conference. Proceedings. Institute of Industrial and Systems Engineers (IISE), 2024, pp. 1–6

    2024

  2. [38]

    A deep reinforcement learning method for solving two-echelon location-routing problem,

    S. Huang, Y . Wu, Z. Cao, and X. Zhang, “A deep reinforcement learning method for solving two-echelon location-routing problem,”Computers & Operations Research, p. 107210, 2025

    2025

  3. [1]

    The effect of ignoring routes when locating depots,

    S. Salhi and G. K. Rand, “The effect of ignoring routes when locating depots,”European Journal of Operational Research, vol. 39, no. 2, pp. 150–156, 1989

    1989

  4. [2]

    Consistency and robustness in location-routing,

    S. Salhi and G. Nagy, “Consistency and robustness in location-routing,” Studies in Locational Analysis, no. 13, pp. 3–19, 1999

    1999

  5. [3]

    The transportation-location problem,

    L. Cooper, “The transportation-location problem,”Operations Research, vol. 20, no. 1, pp. 94–108, 1972

    1972

  6. [4]

    Multi- objective sustainable capacitated location routing problem formulation in sustainable supply-chain management,

    L. F. Galindres, F. G. Guimar ˜aes, and R. A. Gallego-Rend ´on, “Multi- objective sustainable capacitated location routing problem formulation in sustainable supply-chain management,”Engineering Optimization, vol. 55, no. 3, pp. 526–541, 2023

    2023

  7. [5]

    Flexible districting policy for the multiperiod emergency resource allocation problem with demand priority,

    X. Xu, Z. Lin, X. Li, W. Yi, and W. Pedrycz, “Flexible districting policy for the multiperiod emergency resource allocation problem with demand priority,”IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2024

    2024

  8. [6]

    Disaster relief facility network design in metropolises,

    L. Zhen, K. Wang, and H.-C. Liu, “Disaster relief facility network design in metropolises,”IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 45, no. 5, pp. 751–761, 2014

    2014

Show all 48 references
  1. [7]

    A simulated annealing heuristic for the open location-routing problem,

    F. Y . Vincent and S.-Y . Lin, “A simulated annealing heuristic for the open location-routing problem,”Computers & Operations Research, vol. 62, pp. 184–196, 2015

    2015

  2. [8]

    Crowdsourced humanitarian relief vehicle routing problem,

    J. Parappathodi and C. Archetti, “Crowdsourced humanitarian relief vehicle routing problem,”Computers & Operations Research, vol. 148, p. 105963, 2022

    2022

  3. [9]

    Vehicle routing problem and related algorithms for logistics distribution: A literature review and classification,

    G. D. Konstantakopoulos, S. P. Gayialis, and E. P. Kechagias, “Vehicle routing problem and related algorithms for logistics distribution: A literature review and classification,”Operational Research, vol. 22, no. 3, pp. 2033–2062, 2022

  4. [10]

    A survey of recent research on location- routing problems,

    C. Prodhon and C. Prins, “A survey of recent research on location- routing problems,”European Journal of Operational Research, vol. 238, no. 1, pp. 1–17, 2014

    2014

  5. [11]

    An exact method for the capacitated location-routing problem,

    R. Baldacci, A. Mingozzi, and R. Wolfler Calvo, “An exact method for the capacitated location-routing problem,”Operations Research, vol. 59, no. 5, pp. 1284–1296, 2011

    2011

  6. [12]

    Variable neighbor- hood search for location routing,

    B. Jarboui, H. Derbel, S. Hanafi, and N. Mladenovi ´c, “Variable neighbor- hood search for location routing,”Computers & Operations Research, vol. 40, no. 1, pp. 47–57, 2013

    2013

  7. [13]

    Deep reinforcement learning for multi-period facility location: pk-median dynamic location problem,

    C. Miao, Y . Zhang, T. Wu, F. Deng, and C. Chen, “Deep reinforcement learning for multi-period facility location: pk-median dynamic location problem,” inProceedings of the 32nd ACM International Conference on Advances in Geographic Information Systems, 2024, pp. 1–11

    2024

  8. [14]

    Sponet: solve spatial optimization problem using deep reinforcement learning for urban spatial decision analysis,

    H. Liang, S. Wang, H. Li, L. Zhou, H. Chen, X. Zhang, and X. Chen, “Sponet: solve spatial optimization problem using deep reinforcement learning for urban spatial decision analysis,”International Journal of Digital Earth, vol. 17, no. 1, p. 2299211, 2024

    2024

  9. [15]

    Recovnet: Reinforcement learning with covering information for solving maximal coverage billboards location problem,

    Y . Zhong, S. Wang, H. Liang, Z. Wang, X. Zhang, X. Chen, and C. Su, “Recovnet: Reinforcement learning with covering information for solving maximal coverage billboards location problem,”International Journal of Applied Earth Observation and Geoinformation, vol. 128, p. 103710, 2024

    2024

  10. [16]

    Pomo: Policy optimization with multiple optima for reinforcement learning,

    Y .-D. Kwon, J. Choo, B. Kim, I. Yoon, Y . Gwon, and S. Min, “Pomo: Policy optimization with multiple optima for reinforcement learning,”Advances in Neural Information Processing Systems, vol. 33, pp. 21 188–21 198, 2020

    2020

  11. [17]

    Deep reinforcement learning for solving vehicle routing problems with backhauls,

    C. Wang, Z. Cao, Y . Wu, L. Teng, and G. Wu, “Deep reinforcement learning for solving vehicle routing problems with backhauls,”IEEE Transactions on Neural Networks and Learning Systems, 2024

    2024

  12. [18]

    Multi-type attention for solving multi-depot vehicle routing problems,

    J. Li, B. T. Dai, Y . Niu, J. Xiao, and Y . Wu, “Multi-type attention for solving multi-depot vehicle routing problems,”IEEE Transactions on Intelligent Transportation Systems, 2024

    2024

  13. [19]

    A new approach for solving location routing problems with deep reinforcement learning of emergency medical facility,

    S. Wang, J. Zhou, H. Liang, Z. Wang, C. Su, and X. Li, “A new approach for solving location routing problems with deep reinforcement learning of emergency medical facility,” inProceedings of the 8th ACM SIGSPATIAL International Workshop on Security Response using GIS, 2023, pp. 50–53

    2023

  14. [20]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in Neural Information Processing Systems, vol. 30, 2017, pp. 5998– 6008

    2017

  15. [21]

    Branch-and-bound methods: A survey,

    E. L. Lawler and D. E. Wood, “Branch-and-bound methods: A survey,” Operations Research, vol. 14, no. 4, pp. 699–719, 1966

    1966

  16. [22]

    An exact algorithm for minimizing routing and operating costs in depot location,

    G. Laporte and Y . Nobert, “An exact algorithm for minimizing routing and operating costs in depot location,”European Journal of Operational Research, vol. 6, no. 2, pp. 224–226, 1981

    1981

  17. [23]

    Location-routing problems with distance constraints,

    R. T. Berger, C. R. Coullard, and M. S. Daskin, “Location-routing problems with distance constraints,”Transportation Science, vol. 41, no. 1, pp. 29–43, 2007

    2007

  18. [24]

    A branch-and-price algorithm for combined location and routing problems under capacity restrictions,

    Z. Akca, R. Berger, and T. Ralphs, “A branch-and-price algorithm for combined location and routing problems under capacity restrictions,” inOperations Research and Cyber-Infrastructure. Springer, 2009, pp. 309–330

    2009

  19. [25]

    A branch-and-cut method for the capacitated location-routing problem,

    J.-M. Belenguer, E. Benavent, C. Prins, C. Prodhon, and R. W. Calvo, “A branch-and-cut method for the capacitated location-routing problem,” Computers & Operations Research, vol. 38, no. 6, pp. 931–941, 2011

    2011

  20. [26]

    An exact algorithm based on cut-and-column generation for the capacitated location-routing problem,

    C. Contardo, J.-F. Cordeau, and B. Gendron, “An exact algorithm based on cut-and-column generation for the capacitated location-routing problem,”INFORMS Journal on Computing, vol. 26, no. 1, pp. 88–102, 2014

    2014

  21. [27]

    Solving the capacitated location-routing problem by a grasp complemented by a learning process and a path relinking,

    C. Prins, C. Prodhon, and R. W. Calvo, “Solving the capacitated location-routing problem by a grasp complemented by a learning process and a path relinking,”4OR, vol. 4, pp. 221–238, 2006

    2006

  22. [28]

    A grasp×els approach for the capacitated location-routing problem,

    C. Duhamel, P. Lacomme, C. Prins, and C. Prodhon, “A grasp×els approach for the capacitated location-routing problem,”Computers & Operations Research, vol. 37, no. 11, pp. 1912–1923, 2010

    1912

  23. [29]

    An iterated local search for solving a location-routing problem,

    H. Derbel, B. Jarboui, S. Hanafi, and H. Chabchoub, “An iterated local search for solving a location-routing problem,”Electronic Notes in Discrete Mathematics, vol. 36, pp. 875–882, 2010

    2010

  24. [30]

    A simulated annealing heuristic for the capacitated location routing problem,

    F. Y . Vincent, S.-W. Lin, W. Lee, and C.-J. Ting, “A simulated annealing heuristic for the capacitated location routing problem,”Computers & Industrial Engineering, vol. 58, no. 2, pp. 288–299, 2010

    2010

  25. [31]

    Large composite neighborhoods for the capacitated location-routing problem,

    M. Schneider and M. L ¨offler, “Large composite neighborhoods for the capacitated location-routing problem,”Transportation Science, vol. 53, no. 1, pp. 301–318, 2019

    2019

  26. [32]

    Multi- objective multiple neighborhood search algorithms for multiobjective fleet size and mix location-routing problem with time windows,

    J. Wang, L. Yuan, Z. Zhang, S. Gao, Y . Sun, and Y . Zhou, “Multi- objective multiple neighborhood search algorithms for multiobjective fleet size and mix location-routing problem with time windows,”IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 51, no. 4, p...

    2019

  27. [33]

    Pointer networks,

    O. Vinyals, M. Fortunato, and N. Jaitly, “Pointer networks,”Advances in Neural Information Processing Systems, vol. 28, 2015. THIS WORK HAS BEEN SUBMITTED TO THE IEEE FOR POSSIBLE PUBLICATION. COPYRIGHT MAY BE TRANSFERRED WITHOUT NOTICE, AFTER WHICH THIS VERSION MAY NO LONGER ...

    2015

  28. [34]

    Attention, learn to solve routing problems!

    W. Kool, H. van Hoof, and M. Welling, “Attention, learn to solve routing problems!” in7th International Conference on Learning Representa- tions, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019, 2019

    2019

  29. [35]

    Heuristic sequencing hopfield neural network for pick-and-place location routing in multi-functional placers,

    Z. Li, H. Sun, X. Yu, and W. Sun, “Heuristic sequencing hopfield neural network for pick-and-place location routing in multi-functional placers,” Neurocomputing, vol. 472, pp. 35–44, 2022

    2022

  30. [36]

    A reinforcement learning guided hybrid evolutionary algorithm for the latency location routing problem,

    Y . Zou, J.-K. Hao, and Q. Wu, “A reinforcement learning guided hybrid evolutionary algorithm for the latency location routing problem,” Computers & Operations Research, vol. 170, p. 106758, 2024

    2024

  31. [39]

    Simple statistical gradient-following algorithms for connectionist reinforcement learning,

    R. J. Williams, “Simple statistical gradient-following algorithms for connectionist reinforcement learning,”Machine Learning, vol. 8, pp. 229–256, 1992

    1992

  32. [40]

    Deep reinforcement learning for sequence-to-sequence models,

    Y . Keneshloo, T. Shi, N. Ramakrishnan, and C. K. Reddy, “Deep reinforcement learning for sequence-to-sequence models,”IEEE Trans- actions on Neural Networks and Learning Systems, vol. 31, no. 7, pp. 2469–2489, 2019

    2019

  33. [41]

    Empirical evaluation of gated recurrent neural networks on sequence modeling,

    J. Chung, C. Gulcehre, K. Cho, and Y . Bengio, “Empirical evaluation of gated recurrent neural networks on sequence modeling,”arXiv preprint arXiv:1412.3555, 2014

    2014 arXiv

  34. [42]

    Simulation-guided beam search for neural combinatorial optimization,

    J. Choo, Y .-D. Kwon, J. Kim, J. Jae, A. Hottung, K. Tierney, and Y . Gwon, “Simulation-guided beam search for neural combinatorial optimization,”Advances in Neural Information Processing Systems, vol. 35, pp. 8760–8772, 2022

    2022

  35. [43]

    Nouveaux algorithmes pour le probl `eme de localisation et routage avec contraintes de capacit ´e,

    C. Prins, C. Prodhon, and R. W. Calvo, “Nouveaux algorithmes pour le probl `eme de localisation et routage avec contraintes de capacit ´e,” in MOSIM’04 (4 `eme Conf. Francophone de Mod ´elisation et Simulation), 2004

    2004

  36. [44]

    Gurobi Optimizer Reference Manual,

    Gurobi Optimization, LLC, “Gurobi Optimizer Reference Manual,”

  37. [45]

    Two-echelon routing problem for parcel delivery by cooperated truck and drone,

    Y . Liu, Z. Liu, J. Shi, G. Wu, and W. Pedrycz, “Two-echelon routing problem for parcel delivery by cooperated truck and drone,”IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 51, no. 12, pp. 7450–7465, 2020

    2020

  38. [46]

    A hybrid multiobjective memetic algorithm for multiobjective periodic vehicle routing problem with time windows,

    J. Wang, W. Ren, Z. Zhang, H. Huang, and Y . Zhou, “A hybrid multiobjective memetic algorithm for multiobjective periodic vehicle routing problem with time windows,”IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 50, no. 11, pp. 4732–4745, 2018

    2018

  39. [47]

    Vehicle routing problems for drone delivery,

    K. Dorling, J. Heinrichs, G. G. Messier, and S. Magierowski, “Vehicle routing problems for drone delivery,”IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 47, no. 1, pp. 70–85, 2016

    2016

  40. [2023]

    Available: https://www.gurobi.com

    [Online]. Available: https://www.gurobi.com

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.