Learning to Handle Complex Constraints for Vehicle Routing Problems

Vehicle Routing Problems (VRPs) can model many real-world scenarios and often involve complex constraints. While recent neural methods excel in constructing solutions based on feasibility masking, they struggle with handling complex constraints, especially when obtaining the masking itself is NP-hard. In this paper, we propose a novel Proactive Infeasibility Prevention (PIP) framework to advance the capabilities of neural methods towards more complex VRPs. Our PIP integrates the Lagrangian multiplier as a basis to enhance constraint awareness and introduces preventative infeasibility masking to proactively steer the solution construction process. Moreover, we present PIP-D, which employs an auxiliary decoder and two adaptive strategies to learn and predict these tailored masks, potentially enhancing performance while significantly reducing computational costs during training. To verify our PIP designs, we conduct extensive experiments on the highly challenging Traveling Salesman Problem with Time Window (TSPTW), and TSP with Draft Limit (TSPDL) variants under different constraint hardness levels. Notably, our PIP is generic to boost many neural methods, and exhibits both a significant reduction in infeasible rate and a substantial improvement in solution quality.

Paper

References (77)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer Seke5/10 · confidence 4/52024-07-04

Summary

This paper introduces a novel framework called Proactive Infeasibility Prevention (PIP) to enhance the capability of neural methods in addressing complex Vehicle Routing Problems (VRPs) with interdependent constraints, such as the Traveling Salesman Problem with Time Window (TSPTW) and TSP with Draft Limit (TSPDL). The authors propose integrating the Lagrangian multiplier method and preventative infeasibility masking into the solution construction process to improve both the feasibility and optimality of solutions. Additionally, an enhanced version, PIP-D, employs an auxiliary decoder to predict infeasibility masks, potentially reducing computational costs during training. The paper presents extensive experiments demonstrating the effectiveness of PIP in reducing infeasible rates and improving solution quality.

Strengths

The PIP framework represents an innovative approach to integrating constraint awareness and preventative measures within neural VRP solvers. The use of an auxiliary decoder for predicting infeasibility masks is a good try. The paper provides extensive empirical validation demonstrating the effectiveness of the proposed methods. The paper is well-organized and clearly presents the problem, proposed solutions, and experimental results. The use of figures and tables effectively supports the textual content, aiding in the reader's comprehension.

Weaknesses

The ideas of using the Lagrangian multiplier [1] and preventive mask functions[2] are not new. The code is not provided, which limits the reproducibility of the work. It's not clear how to obtain the training label of infeasibility mask. The computation could be expensive to get a global and accurate infeasibility mask. Some related work from recent AI and OR fields are missing. The paper could benefit from additional experiments on a wider range of VRP variants and real-world datasets to further validate the generalizability of the proposed methods. The presentation could be improved by providing more context on how this work relates to and advances the current state-of-the-art in neural VRP solvers. The computational efficiency of the PIP-D framework should be more thoroughly analyzed, especially in terms of how it scales with problem size. The paper may lack a deeper discussion on the theoretical underpinnings of the proposed methods and their potential implications for the broader field of combinatorial optimization. [1] Qiaoyue Tang, Yangzhe Kong, Lemeng Pan, and Choonmeng Lee. Learning to solve soft-constrained vehicle routing problems with lagrangian relaxation. arXiv preprint arXiv:2207.09860, 2022. [2] Hou, Qingchun, et al. Generalize learned heuristics to solve large-scale vehicle routing problems in real-time. The Eleventh International Conference on Learning Representations. 2023.

Questions

How does the PIP framework generalize to other VRP variants beyond TSPTW and TSPDL? Can the authors provide more insights into the computational efficiency of PIP-D, especially regarding its scalability?

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors have acknowledged some limitations, such as the potential inability of PIP to improve performance on all backbone solvers and VRP variants. However, the paper could benefit from a more detailed discussion on the robustness of the PIP framework when faced with different levels of constraint hardness in real-world scenarios.

Reviewer Xb4p5/10 · confidence 4/52024-07-08

Summary

This paper addresses the challenge of predicting feasible solutions for VRPs with complex constraints. It introduces two novel methods to enhance existing algorithms: i) integrating constraints directly into the optimization objective using the Lagrangian Multiplier Approach; ii) PIP framework, which proactively excludes potential nodes during solution prediction to prevent future infeasibility. Besides, a neural decoder PIP-D is proposed to reduce computational complexities. Experiments on two VRP variants demonstrates significant advancements in feasible prediction compared to the baselines.

Strengths

1. The paper is well-structured with a logical flow. 2. The feasibility issue investigated are critical, and the proposed PIP framework address it well. 3. Extensive experiments and detailed clarifications are made to demonstrate effectiveness.

Weaknesses

Major concerns: - **Computational cost.** From the numerical results, PIP incurs approximately twice the computational time compared to baseline neural solvers (e.g., AM and POMO). Despite attempts to mitigate this by introducing neural decoder PIP-D, the reported results don't exhibit a corresponding reduction in time. The authors may clarify this more. - **Lack of comparisons** between different step numbers in PIP. The authors only conducted experiments of one-step PIP. It is crucial to include comparisons with different step numbers to assess the impact of the "looking ahead" masking strategy. For instance, additional results from zero-step PIP (directly masking nodes violating constraints) and two-step PIP. Here, I emphasize zero-step PIP since its masking strategy differs from the ones of the baseline neural solvers, as I understand it. - Only two TSP datasets are considered. The paper is titled "Learning to Handle Complex Constraints for **Vehicle Routing Problems (VRPs)**", however both datasets are **variants of TSPs**. The key difference between VRPs and TSPs is **capacity constraints**. I would like to see VRPs with various complex constraints (such as time windows, pick-up and delivery, split delivery etc.), otherwise, the authors should change the title and consider more TSP variants. - Heuristic algorithms for TSPs and VRPs with complex constraints have been well developed (such as LKH3). I do not see that the proposed algorithm outperforms the state-of-the-art heuristic algorithms (not neural algorithms).

Questions

- What's the time cost of PIP? Is it $time(POMO^*+PIP)$ - $time(POMO^*)$? - What are the masking strategies of baselines AM and POMO? Do they only mask visited nodes, ignoring other constraints? - What is the stopping criteria for each algorithm? I know that running LKH3 for 1 second would often generate almost optimal solutions to TSPs with 100 customers, hence there is no need to run it for 14 hours, as shown in Table 2. I recommend the authors use the same time limit (and usually a few seconds for small-sized TSPs considered in this paper) for each baseline algorithm and even present the progress of the objective value for the best solutions found by each algorithm. This could ensure a fair comparison among them.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

The author has discussed the limitations in section 6: one potential limitation is that it may not improve performance on all backbone solvers and all VRP variants.

Authorsrebuttal2024-08-10

Further results for comprehensive comparison (1/4)

**Table R5**: Results on LKH3 with the similar $\underline{\text {instance inference time}}$ limit as POMO*+PIP(-D). | | | | $n=50$ | | | $n=100$ | | |:------------:|:--------------:|:-------------------------------:|:--------:|:----------------:|:-------------------------------:|:--------:|:----------------:| | **Hardness** | **Method** | $\underline{\text{Inst. Time}}$ | **Obj.** | **Inst. Infsb%** | $\underline{\text{Inst. Time}}$ | **Obj.** | **Inst. Infsb%** | | Easy | LKH3 (Default) | 27s | 7.31 | 0.00% | 49s | 10.21 | 0.00% | | Easy | LKH3 | 0.37s | 7.35 | 0.00% | 0.9s | 10.37 | 0.00% | | Easy | POMO*+PIP | 0.38s | 7.50 | 0.00% | 0.9s | 10.57 | 0.00% | | Easy | POMO*+PIP-D | 0.38s | 7.49 | 0.00% | 0.9s | 10.66 | 0.00% | | | | | | | | | | | Medium | LKH3 (Default) | 40s | 13.02 | 0.00% | 1.0m | 18.74 | 0.00% | | Medium | LKH3 | 0.37s | 13.06 | 0.00% | 0.9s | 19.00 | 0.00% | | Medium | POMO*+PIP | 0.38s | 13.40 | 0.90% | 0.9s | 19.61 | 0.19% | | Medium | POMO*+PIP-D | 0.38s | 13.45 | 0.65% | 0.9s | 19.79 | 0.03% | | | | | | | | | | | Hard | LKH3 (Default) | 40s | 25.61 | 0.12% | 3.2m | 51.24 | 0.07% | | Hard | LKH3 | 0.37s | 25.43 | 30.60% | 0.9s | 49.94 | 97.28% | | Hard | POMO*+PIP | 0.38s | 25.66 | 2.67% | 0.9s | 51.42 | 16.27% | | Hard | POMO*+PIP-D | 0.38s | 25.69 | 3.07% | 0.9s | 51.39 | 6.48% |

Authorsrebuttal2024-08-10

Further results for comprehensive comparison (2/4)

**Table R6**: Results on LKH3 with the similar $\underline{\text {total inference time}}$ limit as POMO*+PIP(-D). | | | | $n=50$ | | | $n=100$ | | |:------------:|:--------------:|:-------------------------------:|:--------:|:----------------:|:-------------------------------:|:--------:|:----------------:| | **Hardness** | **Method** | $\underline{\text{Total Time}}$ | **Obj.** | **Inst. Infsb%** | $\underline{\text{Total Time}}$ | **Obj.** | **Inst. Infsb%** | | Easy | LKH3 (Default) | 4.6h | 7.31 | 0.00% | 8.5h | 10.21 | 0.00% | | Easy | LKH3 | 26s | 8.81 | 99.29% | 58s | / | 100.00% | | Easy | POMO*+PIP | 21s | 7.50 | 0.00% | 48s | 10.57 | 0.00% | | Easy | POMO*+PIP-D | 21s | 7.49 | 0.00% | 48s | 10.66 | 0.00% | | | | | | | | | | | Medium | LKH3 (Default) | 7h | 13.02 | 0.00% | 10.8h | 18.74 | 0.00% | | Medium | LKH3 | 25s | 13.05 | 39.91% | 63s | / | 100.00% | | Medium | POMO*+PIP | 21s | 13.40 | 0.90% | 48s | 19.61 | 0.19% | | Medium | POMO*+PIP-D | 21s | 13.45 | 0.65% | 48s | 19.79 | 0.03% | | | | | | | | | | | Hard | LKH3 (Default) | 7h | 25.61 | 0.12% | 1.4d | 51.24 | 0.07% | | Hard | LKH3 | 22s | / | 100.00% | 54s | / | 100.00% | | Hard | POMO*+PIP | 21s | 25.66 | 2.67% | 48s | 51.42 | 16.27% | | Hard | POMO*+PIP-D | 21s | 25.69 | 3.07% | 48s | 51.39 | 6.48% |

Authorsrebuttal2024-08-10

Further results for comprehensive comparison (3/4)

**Table R7**: Results on POMO*+PIP(-D)+LKH3 with the similar $\underline{\text {instance inference time}}$ limit as doubled time used in POMO*+PIP(-D). | | | | $n=50$ | | | $n=100$ | | |:------------:|:----------------:|:-------------------------------:|:--------:|:----------------:|:-------------------------------:|:--------:|:----------------:| | **Hardness** | **Method** | $\underline{\text{Inst. Time}}$ | **Obj.** | **Inst. Infsb%** | $\underline{\text{Inst. Time}}$ | **Obj.** | **Inst. Infsb%** | | Easy | LKH3 | 0.77s | 7.33 | 0.00% | 1.8s | 10.31 | 0.00% | | Easy | POMO*+PIP+LKH3 | 0.77s | 7.33 | 0.00% | 1.8s | 10.29 | 0.00% | | Easy | POMO*+PIP-D+LKH3 | 0.77s | 7.33 | 0.00% | 1.8s | 10.31 | 0.00% | | | | | | | | | | | Medium | LKH3 | 0.77s | 13.04 | 0.00% | 1.8s | 18.89 | 0.00% | | Medium | POMO*+PIP+LKH3 | 0.77s | 13.05 | 0.00% | 1.8s | 18.91 | 0.00% | | Medium | POMO*+PIP-D+LKH3 | 0.77s | 13.05 | 0.00% | 1.8s | 18.92 | 0.00% | | | | | | | | | | | Hard | LKH3 | 0.76s | 25.56 | 8.81% | 1.8s | 50.66 | 75.54% | | Hard | POMO*+PIP+LKH3 | 0.76s | 25.61 | 0.12% | 1.8s | 51.24 | 1.63% | | Hard | POMO*+PIP-D+LKH3 | 0.76s | 25.61 | 0.05% | 1.8s | 51.25 | 0.36% |

Authorsrebuttal2024-08-10

Further results for comprehensive comparison (4/4)

**Table R8**: Results on POMO*+PIP(-D)+LKH3 with the similar $\underline{\text {total inference time}}$ limit as doubled time used in POMO*+PIP(-D). | | | | | $n=50$ | | | | $n=100$ | | |:------------:|:----------------:|:---------:|:-------------------------------:|:-------------:|:----------------:|:------------:|:-------------------------------:|:---------:|:----------------:| | **Hardness** | **Method** | | $\underline{\text{Total Time}}$ | **Obj.** | **Inst. Infsb%** | | $\underline{\text{Total Time}}$ | **Obj.** | **Inst. Infsb%** | | Easy | LKH3 | | 46s | 7.57 | 0.17% | | 1.8m | 11.17 | 2.54% | | Easy | POMO*+PIP+LKH3 | | 48s | 7.48 | 0.20% | | 1.8m | 10.46 | 0.01% | | Easy | POMO*+PIP-D+LKH3 | | 48s | 7.47 | 0.32% | | 1.8m | 10.52 | 0.00% | | | | | | | | | | | | | Medium | LKH3 | | 46s | 13.36 | 0.67% | | 1.8m | 20.42 | 26.66% | | Medium | POMO*+PIP+LKH3 | | 48s | 13.37 | 1.12% | | 1.9m | 19.38 | 0.20% | | Medium | POMO*+PIP-D+LKH3 | | 48s | 13.41 | 0.77% | | 1.9m | 19.50 | 0.05% | | | | | | | | | | | | | Hard | LKH3 | | 44s | 23.87 | 99.09% | | 1.9m | / | 100.00% | | Hard | POMO*+PIP+LKH3 | | 44s | 25.65 | 2.99% | | 1.8m | 51.39 | 15.80% | | Hard | POMO*+PIP-D+LKH3 | | 44s | 25.68 | 3.23% | | 1.8m | 51.37 | 6.24% | #### #### **Table R9**: Results on LKH3 with the similar $\underline{\text {instance inference time}}$ limit as GFACS*+PIP(-D) (+LKH3) on $n=500$ (first 10 instances). | | | $n=500$ | | | :---------------: | :------------: | :--------------: | :----------: | | **Method** | **Inst. Time** | **Inst. Infsb%** | **Gap to #** | | LKH3 (Default) | 26m | 0.00% | # | | LKH3 | 6.5m | 0.00% | 0.59% | | GFACS* | 6.4m | 57.81% | 21.32% | | GFACS*+PIP-D | 6.5m | 0.00% | 11.52% | | LKH3 | 13m | 0.00% | 0.25% | | GFACS*+PIP-D+LKH3 | 13m | 0.00% | 1.10% |

Reviewer DEtT6/10 · confidence 5/52024-07-09

Summary

The paper propose a novel Proactive Infeasibility Prevention (PIP) framework to advance the capabilities of neural methods towards more complex VRPs, and further investigates the Lagrange multiplier method for soft objective in VRPs, presenting the PIP (& PIP-D) for the hard case where Lagrange multiplier method difficult to find feasible solution. And the experiments show that the PIP is prior to Lagrange multiplier method and has generality to TSPTW and TSPDL.

Strengths

This paper contributes to the solution of the soft constraint objectives of VRPs. The author's perspective on the problem is also good; that is, the constraint itself is an NP-hard problem, which was not mentioned in reference [8], and the description of why the constraint itself is NP-hard is very clear, that is, because of the irreversible impact of the first selected node on the selection of subsequent customer nodes.

Weaknesses

The author's method actually addresses the shortcomings of the Lagrange multiplier method, but it is not mentioned in the Introduction, and the problems raised in the Introduction can actually be solved only using the Lagrange multiplier method. The method involves complex calculations, particularly when integrating the Lagrangian multiplier and PI masking, which can be computationally intensive​. The paper does not thoroughly address the scalability of the proposed method when applied to extremely large datasets, which may limit its practical application in some real-world scenarios​​.

Questions

[1]What are the potential challenges and limitations in scaling the PIP method to larger and more complex combinatorial optimization problems? [2]How can we further optimize or modify the proposed method to reduce computational complexity without sacrificing solution quality or feasibility?

Rating

6

Confidence

5

Soundness

3

Presentation

2

Contribution

3

Limitations

yes

Area Chair JCXx2024-08-12

Dear Reviewers DEtT and Seke:

The authors have provided extensive comments and new results in response to the criticisms raised in your reviews. Has this response addressed your main concerns? If not, are there additional questions you would like to pose? The author discussion period is scheduled to end tomorrow (Aug 15). Please respond.

Reviewer azSu7/10 · confidence 4/52024-07-12

Summary

This paper proposes a Proactive Infeasibility Prevention (PIP) framework to enhance the ability of neural methods to handle complex constraints in Vehicle Routing Problem (VRP). The PIP framework integrates the Lagrangian multiplier to enhance constraint awareness and introduces preventative infeasibility masking to proactively guide the solution construction process. Additionally, an extended version called PIP-D employs an auxiliary decoder and two adaptive strategies to learn and predict masking information, reducing computational costs during training. These methods were extensively tested on different levels of constraint hardness in the Traveling Salesman Problem with Time Window (TSPTW) and Traveling Salesman Problem with Draft Limit (TSPDL) variants. The results demonstrate that the proposed methods enhance the capabilities of neural methods, significantly reducing infeasibility rates and improving solution quality.

Strengths

1. Adaptability to Complex Constraints: The paper devises a method to apply machine learning to constraint-based problems, which have been challenging to handle with deep learning due to the difficulty in finding feasible solutions. 2. Experimental Validation: Experimental results demonstrate that the proposed method can compute feasible solutions more effectively compared to traditional methods. 3. Integration with Constructive Methods: The proposed approach can be combined with constructive methods, enhancing its practical applicability and flexibility.

Weaknesses

1. Generality of the Trained Model: It is unclear whether the trained model generalizes well to unseen instances, raising concerns about its robustness and applicability to different problem settings. 2. Lack of Theoretical Justification for PIP: The paper does not provide a theoretical justification for the superiority of using the Proactive Infeasibility Prevention (PIP) method, leaving its theoretical advantages unproven. 3. Lack of Classical methods with time limit: There are no comparison with the classical methods (e.g., LKH3) with a time limit. I am concerned that LKH3 could find "good" feasible solutions within a short time (e.g., within 5 minutes).

Questions

- Can the proposed idea be applied to methods other than constructive methods? If it can be used to guide the search to avoid infeasibility, it would be a more practical idea. - Alternatively, after the (possibly infeasible) solution constructed by a constructive method, is it easy to recover a feasibility through neighborhood search? If so, can you compare this with the avoidance of infeasibility using PIP? - Is the model constructed with PIP tailored to the training dataset? In other words, when using the model constructed with PIP for instances in different area, does the number of infeasible instances decrease compared to the model without PIP? - Is it possible to use the PIP framework for more general VRP (e.g., with multiple vehicles or capacity constraints)? - Are there any experimental comparisons between the proposed model and LKH3 with time limit?

Rating

7

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

The limitations of this work are discussed in the last section.

Reviewer Xb4p2024-08-08

Thanks for addressing my comments. I still think the comparison between the proposed approach and LKH3 is not comprehensive enough (e.g., easy/medium/hard instances with more than 100 nodes, and the 500-customer instances in Section 5.2; **with the same time limit as the proposed approach**) since **LKH3 is the most important baseline algorithm** in my opinion. For example, from Figure R3, we may conclude that, for those hard instances, the proposed approach could obtain high-quality solutions quite fast but LKH3 could obtain solutions of better quality with a relatively longer time. But from Section 5.2, if we run LKH3 for a few minutes (say 6.5 min), we may already achieve a very small gap (e.g., smaller than 11.95%). That is why I said the comparison in the previous manuscript is not fair.

Authorsrebuttal2024-08-10

Thanks for the comments and please kindly check our further clarifications.

Thank you for the discussion and insightful comments! We are pleased to have addressed most of your concerns. Regarding the remaining concern about a comprehensive comparison between our PIP and LKH3, we have provided results on Hard TSPTW-100 in the previous rebuttal given the constraints of the rebuttal period and PDF page limit. Nevertheless, we understand that this may not be comprehensive enough. Following your suggestion, we now conducted additional experiments using LKH3 with the same time limits as our proposed approach **across various instance hardness levels (Easy/Medium/Hard) and scales (50/100/500 nodes)**. The time limits were set in two ways for comprehensive evaluation: similar $\underline{\text {instance inference time}}$ without parallelization (new Tables R5, R7, and R9) and similar $\underline{\text {total inference time}}$ with parallelization (new Tables R6 and R8). We hope these new empirical results offer a more comprehensive picture. **[A. For $n=50$ and $100$]** Firstly, we follow your suggestion and only allow LKH3 to run a similar inference time as our approach $\underline{\text {per instance}}$. As shown in **Table R5**, POMO*+PIP(-D) outperforms LKH3 on Hard datasets, while maintaining competitive results on Easy and Medium datasets. While comparing per-instance time might seem fair for CPU-based LKH3, ignoring parallelization could disadvantage GPU-based solvers (thus not fair for GPU-based solvers). To further explore this, we conducted another experiment but with the similar $\underline{\text {total inference time}}$ limit across both methods. Results, in **Table R6**, show that our POMO*+PIP(-D) performs consistently better than LKH3 across most of the hardness. We recognize the challenge of achieving absolute fairness when comparing CPU-based and GPU-based solvers, and we note that total time comparison is a common practice in most existing NCO papers (e.g. those in [1-8]). We appreciate the reviewer for providing such an insightful comment, which highlights the need for developing new metrics that facilitate fairer comparisons between neural and traditional solvers—a promising direction for future work. Moreover, we agree with the reviewer that LKH3 is indeed a robust and well-developed solver. However, neural solvers offer unique advantages, such as highly efficient parallelization and reduced reliance on hand-crafted heuristic rules [9]. To leverage the strengths of both approaches and further reveal the practical usage of our method, we explored a hybrid method that combines our PIP(-D) framework with LKH3. Similarly, we provide the comprehensive experimental results conducted across various instance hardness levels (Easy/Medium/Hard) and scales (50/100/500 nodes). The results, as shown in **Table R7 and R8**, reveal that **LKH3’s search efficiency can be significantly enhanced when initialized with solutions from our PIP(-D) framework**. **[B. For $n=500$]** As shown in **Table R9**, we acknowledge that our PIP(-D) implemented on GFACS* currently underperforms LKH3. However, the integration of GFACS*+PIP(-D) with LKH3 has already made significant strides in closing this performance gap, particularly in enhancing feasibility handling compared to using the GFACS backbone alone. We would like to clarify that the primary motivation for these experiments here is to demonstrate the scalability and broad applicability of our PIP framework in enhancing various backbone neural models. While we acknowledge that PIP may not elevate every model to state-of-the-art performance, which is both reasonable due to _“No Free Lunch”_, our goal is to show its potential to significantly boost performance across a wide range of neural models.

Authorsrebuttal2024-08-10

Thanks for the comments and please kindly check our further clarifications (continued)

Additionally, we would like to clarify that: 1. **The motivation of this work is to highlight the issue of “NP-hard masking” in neural solvers and enhance neural solvers with the ability to handle complex constraints**, as comprehensively demonstrated in our main paper (Tables 1, 2, and 3). Thus, outperforming LKH3 is not our primary objective or key contribution. Instead, our proposed framework has significantly reduced the infeasibility rate and substantially improved solution quality for both autoregressive and non-autoregressive neural solvers. These advancements establish our framework as the new state-of-the-art in the domain of neural solvers. 2. **Most existing NCO solvers [2-8] use LKH3 with the default settings from [1]** (including maximum trials, number of runs, and parallelism settings) to estimate the optimality gap and provide a runtime reference. The primary focus is on understanding the performance gap between neural solvers [10, 11] and traditional solvers, rather than strictly outperforming it. 3. We acknowledge that **neural solvers have yet to fully replace traditional solvers like LKH3, but rather catching up with it step by step**. We believe that we provide insights for complex constraints handling for NCO, which is a step towards practical applications. Also, we would like to emphasize that our PIP(-D) is a **generic framework**, making our research orthogonal to the development of new, powerful neural solvers that have the potential to fully outperform LKH in future works. We appreciate the reviewer for helping our work to be more rigorous. The new results will be incorporated into the revised manuscript, along with a deeper discussion of these points. Due to the rebuttal limit, we will plot the progress curves (like Figure R2 and R3 in the attached PDF) based on the results of all the new tables for a more intuitive illustration. Also, **our code has already been forwarded to the AC**. We hope our discussions are now comprehensive. Please let us know if you have any remaining concerns. #### #### ``` References: [1] Attention, learn to solve routing problems!. ICLR, 2018. [2] Reinforcement learning for solving the vehicle routing problem. NeurIPS, 2018. [3] POMO: Policy optimization with multiple optima for reinforcement learning. NeurIPS, 2022. [4] Simulation-guided beam search for neural combinatorial optimization. NeurIPS, 2022. [5] Learning to search feasible and infeasible regions of routing problems with flexible neural k-opt. NeurIPS, 2023. [6] BQ-NCO: Bisimulation quotienting for generalizable neural combinatorial optimization. NeurIPS, 2023. [7] Neural combinatorial optimization with heavy decoder: Toward large scale generalization. NeurIPS, 2023. [8] MVMoE: Multi-Task Vehicle Routing Solver with Mixture-of-Experts. ICLM, 2024. [9] Machine learning for combinatorial optimization: a methodological tour d’horizon. European Journal of Operational Research, 2021. [10] Graph-based diffusion solvers for combinatorial optimization. NeurIPS, 2023. [11] MOCO: A Learnable Meta Optimizer for Combinatorial Optimization. arXiv preprint, 2024. ``` #### Note for Tables R7 & R8: For a fair comparison, we kept the POMO*+PIP(-D) models (with inference time denoted as $T_{PIP}$) unchanged and allowed LKH3 to search for a similar duration as $T_{PIP}$ to showcase the performance of POMO*+PIP(-D)+LKH3 (i.e., the total inference time is $2 \times T_{PIP}$). LKH3 was also given the same time limit for consistency.

Reviewer Xb4p2024-08-10

Thanks for addressing my comments. Now I'm satisfied with the comparison and raise my score. I expect the authors to include those results in the updated manuscript and edit part of their conclusions (the proposed approach vs LKH).

Authorsrebuttal2024-08-10

Thank you for your support and we are happy to have addressed your concerns!

We really appreciate your insightful suggestions and we will include the additional results in the revised paper. Thank you!

Authorsrebuttal2024-08-12

Dear Reviewer azSu, Thank you once again for your insightful comments and helpful suggestions. As the deadline for author-reviewer discussions is approaching, we would greatly appreciate it if you could take a moment to review our rebuttal. Please let us know if you have any further questions or concerns. Thank you very much for your time. Best, Authors

Reviewer azSu2024-08-12

Official Comment by Reviewer azSu

Thank you very much for your thoughtful response and the additional analysis. Most of my concerns have been resolved. I understand that the theoretical analysis of PIP and its application to iterative methods are considered future work. I agree that the proposed methods and its experiments presented in this paper are valuable, and I raised my score.

Authorsrebuttal2024-08-12

Dear Reviewer DEtT, Thank you once again for your insightful comments and helpful suggestions. As the deadline for author-reviewer discussions is approaching, we would greatly appreciate it if you could take a moment to review our rebuttal. Please let us know if you have any further questions or concerns. Thank you very much for your time. Best, Authors

Reviewer DEtT2024-08-13

Thank you very much for your thoughtful response and the additional analysis. Most of my concerns have been resolved, and my score has been raised.

Authorsrebuttal2024-08-13

Thank you for your support!

We greatly appreciate your time in reviewing our paper and reading the follow-up rebuttal! We're thrilled for your recognition of our work, and will include the additional analysis in the revised paper. Thanks a lot!

Authorsrebuttal2024-08-12

Dear Reviewer Seke, Thank you once again for your insightful comments and helpful suggestions. As the deadline for author-reviewer discussions is approaching, we would greatly appreciate it if you could take a moment to review our rebuttal. Please let us know if you have any further questions or concerns. Thank you very much for your time. Best, Authors

Authorsrebuttal2024-08-14

Dear Reviewer Seke, #### The discussion period is approaching its end (in less than 12 hours). Please kindly let us know if our response resolves your concerns. Regarding your concern of our code release, we have also forwarded our code to the Area Chair. We would greatly appreciate it if you could give us any feedback. #### Thank you again for your valuable comments and suggestions. #### Best Regards, Authors

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC