acceptodds
Under review as a conference paper at ICLR 2027

Evolving Neural Solvers for Multi-objective Combinatorial Optimization

Abstract

Multi-objective combinatorial optimization (MOCO) seeks high-quality solutions that represent different trade-offs among conflicting objectives. Existing neural methods typically train a dedicated model for each preference or condition a single model on different preferences. Training dedicated models becomes more costly as the number of preferences increases, while a shared model can produce duplicate solutions for different preferences, limiting the benefit of denser preference queries. We propose the Evolving Neural Solver (ENS), which trains only a small set of preference-specific models and takes advantage of evolutionary computation to evolve them into a diverse collection of solvers for better solution quality and Pareto front coverage. ENS first trains a model at an estimated knee-point preference to target a balanced compromise, then obtains the remaining initial models through parameter transfer toward the extremes, so as to preliminarily approximate the Pareto front. It then evolves these models to find more candidates that can generate non-dominated solutions through evolution strategies. Experimental results on classic MOCO problems show that ENS outperforms state-of-the-art neural MOCO methods in terms of hypervolume in most evaluated settings.

Primary area:
optimization
Keywords:
Multi-objective combinatorial optimization, Evolutionary computation, Neural combinatorial optimization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_yZjvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Evolving Neural Solver (ENS), a framework combining reinforcement learning and evolutionary search for multi-objective combinatorial optimization. ENS trains a small set of neural solvers through knee-point initialization and parameter transfer, then evolves their decoder query projections and retains solvers contributing non-dominated solutions. Experiments demonstrate competitive hypervolume results across several MOCO problems, particularly on tri-objective TSP.

Critical strengths
  1. Combining neural solvers with evolutionary search effectively improves Pareto-front coverage.
  2. Restricting evolution to the decoder query projection is a simple and reasonable design choice, supported by the ablation results.
  3. The experiments cover multiple MOCO problems and include strong baselines such as WE-CA, with particularly promising results on tri-objective TSP.
Critical weaknesses
  1. Limited methodological novelty. Neural parameter evolution, parameter transfer, and non-dominated selection have been explored in prior work. The contribution mainly lies in combining existing techniques, without direct comparisons against earlier neural evolutionary MOCO methods.
  2. Unclear inference budgets and comparison fairness. The paper does not report the final number of retained solvers or clearly specify the number of preferences used by neural baselines at inference. Differences in the number of generated solutions may affect HV, making it difficult to attribute the improvements to the proposed method.
  3. Insufficient computational cost analysis. The evolutionary stage evaluates a large number of candidate solvers, but its total runtime and computational cost are not reported. The implementation details of multi-solver inference are also unclear, limiting the assessment of efficiency relative to WE-CA.
  4. Incomplete ablation of key components. The ablations mainly examine the effectiveness of evolving W_q, without isolating the contributions of reseed mutation, elite mutation, direction updates, and global non-dominated selection. It remains unclear whether simpler random perturbations could achieve similar results.
Questions for authors
  1. What are the key advantages of ENS over prior neural evolutionary MOCO methods under comparable computational budgets?
  2. How many solvers are retained after evolution, and how many preferences are evaluated by baselines such as WE-CA? Can the authors provide HV comparisons under matched inference budgets or numbers of generated solutions?
  3. What is the total computational cost of evolution? How are multiple solvers executed at inference, and are encoder computations shared?
  4. How much do direction-guided evolution and global non-dominated selection contribute compared with random perturbations and conventional population selection methods?
3 comments
AuthorsOpenReview ↗

Response to Reviewer yZjv (Part 1/2)

We thank the reviewer for the constructive comments. We address each concern below.

Q1. Advantages over prior neural evolutionary MOCO methods

ENS builds on prior work that evolves neural solvers for MOCO. MODRL/D-EL (Zhang et al., 2021) uses reinforcement learning and parameter transfer to train one model per scalarized subproblem, then evolves the models within a fixed-size population. MONEADD (Shao et al., 2021) evolves a population of randomly initialized networks through crossover and mutation. ENS differs in three ways.

Expanding a small set of trained anchors. MODRL/D-EL first trains or adapts one model for each scalarized subproblem using reinforcement learning and parameter transfer, then evolves these models while keeping the population size fixed. MONEADD starts from randomly initialized networks and evolves the population through crossover and mutation. ENS instead trains an anchor at the estimated knee-point preference and transfers its parameters outward to a few anchors covering different trade-off regions. On Bi-TSP50, only five anchors are trained with reinforcement learning. Evolution then expands these trained starting points into a larger solver ensemble without additional reinforcement learning for each new solver. This decouples the final ensemble size from the number of models trained with reinforcement learning.

Targeted evolution with shared encoding. ENS evolves only the decoder query projection WqW_q. This changes how the decoder attends to candidate nodes while preserving the representations learned by the anchor. It also allows all solvers derived from the same anchor to reuse its encoder outputs. During evolution, perturbing the original anchors sustains exploration, while direction-guided mutations of surviving solvers build on promising search directions.

Selection based on each solver’s contribution to the solution set. MODRL/D-EL already uses non-dominated selection, assigning one fitness vector to each model. ENS evaluates each solver through multiple representative objective vectors, then pools these vectors across anchors. A solver can survive if even one of its representative points is globally non-dominated. This lets solvers that are particularly effective in a specific trade-off region contribute to the final ensemble.

MODRL/D-EL and MONEADD evolve network weights, whereas ENS restricts the search to the decoder query projection WqW_q. This gives evolution a smaller parameter space and targets mutations at the attention mechanism that guides solution construction. ENS also reuses the same rollout trajectories to evaluate each candidate under multiple preferences. Anchor-based mutations continue exploring around trained solvers, while direction-guided mutations refine candidates that have already contributed non-dominated points. Together, these choices make more efficient use of candidate generation and evaluation during evolution.

AuthorsOpenReview ↗

Response to Reviewer yZjv (Part 2/2)

Q2 & Q3. Solver Counts, Inference Budgets, and Evolutionary Cost

Single-model methods. Each method queries one trained model with 101 preferences for bi-objective problems and 105 preferences for tri-objective problems. The preferences are generated using the Das–Dennis simplex-lattice scheme. For two objectives, they are uniformly spaced along the preference segment, including both endpoints:

\qquad i=0,\ldots,100.$$ For three objectives, the preference vectors form a lattice on the two-dimensional simplex: $$\lambda_{i,j,k} =\left(\frac{i}{13},\,\frac{j}{13},\,\frac{k}{13}\right), \qquad i,j,k\in\mathbb{Z}_{\geq0},\quad i+j+k=13.$$ This gives $\binom{15}{2}=105$ preferences, including the simplex vertices and boundary points. **Multi-model methods.** Each method trains or fine-tunes **101 preference-specific models for bi-objective problems** and **105 preference-specific models for tri-objective problems**, with one model corresponding to each preference in the same grids above. Each model is then evaluated on the test instances to produce solutions for its corresponding preference. The number of finally saved models of ENS is variable across problems of different scales. For example, ENS retained 660 solvers in our original Bi-TSP50 run. To make the comparison fairer, we ran additional experiments on Bi-TSP50 using the strongest single-model and multi-model baselines, WE-CA and EMNH, respectively. We increased the number of preference queries for WE-CA from 101 to 701 using the same trained model, and increased the number of preference-specific models for EMNH from 101 to 701. The table below reports HV, inference time, and total offline computation time for all three methods on Bi-TSP50. The reported offline time includes anchor training and evolutionary extension for ENS, 200 training epochs using the original paper’s hyperparameters for WE-CA, and estimated meta-training plus measured fine-tuning for EMNH. Bi-TSP50 with 64-fold augmentation | Method | Queries | Models | Initial training | Fine-tune / evolve | Offline total | HV ↑ | Inference ↓ | |:---|:---:|:---:|:---:|:---:|:---:|:---:|:---:| | **WE-CA** | 101 | 1 | 8.75 h | — | 8.75 h | 0.6412 | 3.3 m | | | 201 | 1 | 8.75 h | — | 8.75 h | 0.6418 | 5.23 m | | | 301 | 1 | 8.75 h | — | 8.75 h | 0.6420 | 7.82 m | | | 401 | 1 | 8.75 h | — | 8.75 h | 0.6421 | 10.41 m | | | 501 | 1 | 8.75 h | — | 8.75 h | 0.6421 | 13.00 m | | | 601 | 1 | 8.75 h | — | 8.75 h | 0.6422 | 15.60 m | | | 701 | 1 | 8.75 h | — | 8.75 h | 0.6422 | 18.25 m | | **EMNH** | — | 101 | 9.58 h | 7.55 m | 9.71 h | 0.6408 | 1.7 m | | | — | 201 | 9.58 h | 10.41 m | 9.76 h | 0.6421 | 4.91 m | | | — | 301 | 9.58 h | 13.42 m | 9.81 h | 0.6427 | 7.35 m | | | — | 401 | 9.58 h | 16.81 m | 9.86 h | 0.6429 | 9.80 m | | | — | 501 | 9.58 h | 20.80 m | 9.93 h | 0.6431 | 12.24 m | | | — | 601 | 9.58 h | 23.15 m | 9.97 h | 0.6433 | 14.69 m | | | — | 701 | 9.58 h | 26.43 m | 10.02 h | 0.6434 | 17.14 m | | **ENS** | — | 660 | 5.00 h | 56.23 m | **5.94 h** | **0.6435** | **7.5 m** | Increasing WE-CA from 101 to 701 queries raises HV from 0.6412 to 0.6422, while inference time grows from 3.3 to 18.25 m. EMNH benefits more from additional models, reaching an HV of 0.6434 with 701 solvers, close to ENS. At comparable query or solver counts, ENS generates solutions faster than both baselines, taking 7.5 m. Solvers derived from the same anchor share its encoder, so the encoder outputs are computed once per input and augmentation view and reused across their decoders. This reuse helps explain ENS’s inference speed at this scale. **Q4. Ablation of Direction Guidance and Joint Non-Dominated Selection** We ran additional ablations on Bi-TSP50 with the same five anchors, random seed, and budget of 100 generations for each variant. **Without direction estimation** removes the estimated directional offset from mutations. **Without reseeding** uses the anchors to initialize the population, then generates offspring from the existing parent pool. **Independent selection within each anchor** applies diversity-based survivor selection separately to each anchor’s population, so candidates from different anchors do not compete. **Reseeding only** generates all offspring by applying independent Gaussian perturbations to the anchors’ query projections. It retains joint survivor selection and the archive, while disabling directional guidance and the use of archived elites as parents. | Evolutionary variant | HV without augmentation ↑ | HV with augmentation ↑ | |:---|:---:|:---:| | **Full ENS** | **0.6400** | **0.6437** | | Without direction estimation | 0.6391 | 0.6435 | | Without reseeding | 0.6390 | 0.6436 | | Independent selection within each anchor | 0.6388 | 0.6427 | | Reseeding only | 0.6341 | 0.6407 | Full ENS achieves the highest HV with and without augmentation.

Thank you for the detailed rebuttal and additional experiments. The additional details on inference budgets and computational costs, together with the new ablation results, address several of my earlier concerns. However, I remain unconvinced by the justification for evolving only WqW_q. The provided derivation merely shows that perturbing (W_q) changes attention scores, which is also true for other projections such as WkW_k. It does not sufficiently explain why WqW_q is a particularly suitable subspace for evolutionary search in MOCO.

Moreover, the additional ablations on Bi-TSP50 show only marginal gains from direction estimation and reseeding, especially with augmentation. While the overall framework is empirically effective, I still find its core design largely heuristic, with limited methodological novelty over existing neural evolutionary approaches. The rebuttal improves the experimental clarity but does not substantially change my assessment of the contribution. I therefore maintain my original score.

Reviewer_ZXgHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Evolving Neural Solvers (ENS), a two-stage framework for multi-objective combinatorial optimization. In the first stage, ENS trains an anchor at an estimated "knee-point" preference and obtains remaining anchors through continuous parameter transfer toward extreme preferences. In the second stage, each anchor serves as a starting point for a population, evolving, being evaluated, and undergoing inference testing within the low-dimensional subspace corresponding to its respective decoder. The method is evaluated on bi-objective and tri-objective variants of the TSP, CVRP, and knapsack problems. The reported results show that ENS achieves the best hypervolume (HV) in 10 out of 12 problem-size combinations.

Critical strengths

The paper presents a comprehensive framework that combines preference-specific anchors with evolutionary search to construct diverse neural solvers for multi-objective combinatorial optimization. Its evaluation spans several problem families, including TSP, CVRP, and knapsack variants, and compares against both neural and conventional optimization baselines, demonstrating that the proposed framework is applicable across multiple problem settings.

Critical weaknesses

1.The experiment result is insufficient to establish robust improvements over strong baselines. The reported hypervolume (HV) gains over competitive baselines are marginal. For example, on Bi-TSP50, ENS-Aug achieves an HV of 0.6435 compared with 0.6412 for WE-CA-Aug, while on Bi-CVRP50, ENS-Aug achieves 0.4109 compared with 0.4106 for EMNH-Aug. More importantly, the paper does not report averaged results over multiple independent runs, standard deviations, confidence intervals, or statistical significance tests. This is particularly concerning given the stochastic nature of neural training and evolutionary search. Without repeated-run evaluation, it is difficult to determine whether the small performance differences are reproducible or attributable to random variation. Consequently, the experimental results do not convincingly establish that ENS provides reliable improvements over existing methods, weakening the empirical basis for the paper's main claims.

2.The efficiency–quality trade-off is not sufficiently demonstrated. A key motivation of ENS is to obtain diverse, high-quality preference-specific solvers without incurring the cost of training a separate model for every preference. However, the evolutionary stage introduces additional candidate generation and evaluation, while the reported quality improvements over strong baselines are small. Although the paper reports runtime comparisons and ablations of different evolved parameter projections, it does not sufficiently establish the end-to-end cost of anchor training and evolutionary search relative to the resulting performance gains, nor does it directly compare the proposed evolutionary strategy with alternative search strategies under comparable computational budgets. As a result, it remains unclear whether ENS offers a meaningful practical advantage over existing approaches or whether its reported gains justify the additional computational effort. This limits the evidence for both the methodological contribution and the practical value of the proposed framework.

Questions for authors

Please see the Weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer ZXgH

We address each concern below.

1. Magnitude and reproducibility of the HV gains

As defined in Appendix B, normalized hypervolume (HV) is the fraction of the reference box dominated by the returned solution set. A higher HV therefore means that the solutions collectively dominate a larger portion of the objective space. ENS achieves appreciable improvements over state-of-the-art baselines on bi-objective problems and ENS-Aug is also the only neural method in Table 1 to outperform WS-LKH on Tri-TSP.

We repeated the evolutionary stage ten times on Bi-TSP50 with independent random seeds, using the same five trained anchors and the same 200 test instances. With augmentation, the mean HV was 0.6436, the sample standard deviation was 2.82×10−52.82 \times 10^{-5}, and the 95% Student's tt confidence interval had a half-width of 2.02×10−52.02 \times 10^{-5}. Without augmentation, the mean HV was 0.6399, the sample standard deviation was 6.40×10−56.40 \times 10^{-5}, and the 95% confidence interval had a half-width of 4.58×10−54.58 \times 10^{-5}. These intervals describe uncertainty in the mean HV across evolutionary runs with fixed anchors and test instances.

2. Quality–Cost Trade-off and Comparison with Alternative Search Strategies

Each individual ENS solver uses the standard POMO architecture. The longer inference time reported for ENS in some settings comes primarily from evaluating more solvers. For example, ENS retained 660 solvers in our original Bi-TSP50 run. To make the comparison fairer, we ran additional experiments on Bi-TSP50 using the strongest single-model and multi-model baselines, WE-CA and EMNH, respectively. We increased the number of preference queries for WE-CA from 101 to 701 using the same trained model, and increased the number of preference-specific models for EMNH from 101 to 701. The table below reports HV, inference time, and total offline computation time for all three methods on Bi-TSP50. The reported offline time includes anchor training and evolutionary extension for ENS, 200 training epochs using the original paper’s hyperparameters for WE-CA, and estimated meta-training plus measured fine-tuning for EMNH.

Bi-TSP50 with 64-fold augmentation

MethodQueriesModelsInitial trainingFine-tune / evolveOffline totalHV ↑Inference ↓
WE-CA10118.75 h—8.75 h0.64123.3 m
20118.75 h—8.75 h0.64185.23 m
30118.75 h—8.75 h0.64207.82 m
40118.75 h—8.75 h0.642110.41 m
50118.75 h—8.75 h0.642113.00 m
60118.75 h—8.75 h0.642215.60 m
70118.75 h—8.75 h0.642218.25 m
EMNH—1019.58 h7.55 m9.71 h0.64081.7 m
—2019.58 h10.41 m9.76 h0.64214.91 m
—3019.58 h13.42 m9.81 h0.64277.35 m
—4019.58 h16.81 m9.86 h0.64299.80 m
—5019.58 h20.80 m9.93 h0.643112.24 m
—6019.58 h23.15 m9.97 h0.643314.69 m
—7019.58 h26.43 m10.02 h0.643417.14 m
ENS—6605.00 h56.23 m5.94 h0.64357.5 m

Increasing WE-CA from 101 to 701 queries raises HV from 0.6412 to 0.6422, while inference time grows from 3.3 to 18.25 m. EMNH benefits more from additional models, reaching an HV of 0.6434 with 701 solvers, close to ENS. ENS has the lowest Inference time at 7.5min and also the lowest total offline time at 5.94 h, compared with 8.75 h for WE-CA and an estimated 9.71–10.02 h for EMNH. Evolutionary search accounts for about 56 m of ENS’s offline time and expands five trained anchors into an ensemble that achieves the highest HV in this comparison.

We ran additional ablations on Bi-TSP50 with the same five anchors, random seed, and budget of 100 generations for each variant. Without direction estimation removes the estimated directional offset from mutations. Without reseeding uses the anchors to initialize the population, then generates offspring from the existing parent pool. Independent selection within each anchor applies diversity-based survivor selection separately to each anchor’s population, so candidates from different anchors do not compete. Reseeding only generates all offspring by applying independent Gaussian perturbations to the anchors’ query projections. It retains joint survivor selection and the archive, while disabling directional guidance and the use of archived elites as parents.

Evolutionary variantHV without augmentation ↑HV with augmentation ↑
Full ENS0.64000.6437
Without direction estimation0.63910.6435
Without reseeding0.63900.6436
Independent selection within each anchor0.63880.6427
Reseeding only0.63410.6407

Full ENS achieves the highest HV with and without augmentation.

Reviewer_ZXgHOpenReview ↗

Official Comment by Reviewer ZXgH

Thank you to the authors for their rebuttal. The additional quality–cost trade-off experiments partly alleviate my original concerns. However, similar as the points from reviewer yZjv, the reported marginal gains are insufficient to support the initial claim with robust improvements over strong baselines. Therefore, I decide to maintain my original score.

Reviewer_Sz9GOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

In this paper, the authors propose a novel method, Evolving Neural Solver (ENS), for solving multi-objective combinatorial optimization (MOCO) problems. The motivation for the proposed framework is that existing methods either require training dedicated models for different preferences, which can be costly, or use a single model conditioned on different preferences, which may produce duplicate solutions across preferences. ENS consists of two main stages: (1) Knee-point initialization and transfer, which creates a set of anchor models starting from an anchor corresponding to the estimated knee-point preference of the Pareto front, and (2) Evolutionary extension, which uses evolutionary search to expand the trained anchor models into a diverse collection of neural solvers. The paper provides ablation studies of some components of the framework and empirical comparisons with relevant baselines, demonstrating improvements in most evaluated settings.

Critical strengths
  • The paper considers an important problem of multi-objective combinatorial optimization (MOCO).
  • The paper evaluates the proposed framework on multiple MOCO benchmarks and compares its performance against a wide range of neural and conventional optimization baselines.
Critical weaknesses
  • The proposed framework shows relatively marginal improvements over existing baselines in many settings, despite introducing additional computational complexity. In some settings, ENS also requires substantially higher inference time than the baselines, particularly when instance augmentation is used.
  • It is difficult to assess the significance of the reported improvements over existing baselines, as the results are often very close, and no confidence intervals or standard deviations are provided.
  • The ablation studies do not fully illustrate the contribution of each component of the proposed framework. While the evolutionary stage provides improvements over using anchors alone, the benefits of individual design choices appear marginal in Table 2. Furthermore, several components, such as the different mutation strategies and search direction updates, are not separately evaluated, making it difficult to justify the additional complexity introduced by the framework.
  • While the proposed framework aims to reduce the cost of training multiple preference-specific models, the evolutionary stage introduces additional computational overhead through the evaluation of a large number of candidate solvers. The paper does not provide a detailed comparison of the total training and evolutionary costs against the baselines.
  • The table captions in the paper could be improved by providing more details about the metrics used and the context of the results presented in each table.
Questions for authors
  • Given the relatively marginal improvements over existing baselines in many settings, can the authors further justify the additional computational complexity introduced by the proposed framework?
  • Can the authors provide standard deviations or confidence intervals for the reported results to better assess the significance of the improvements over existing baselines?
  • Can the authors provide more comprehensive ablation studies to illustrate the contribution of individual components of ENS?
  • Can the authors provide a comparison of the total computational cost, including anchor training and evolutionary search, against the existing baselines?
3 comments
AuthorsOpenReview ↗

Response to Reviewer Sz9G (Part 1/2)

We thank the reviewer for the constructive comments. We address each concern below.

Q1 & Q4. Solution quality and computational cost

As defined in Appendix B, normalized hypervolume (HV) is the fraction of the reference box dominated by the returned solution set. A higher HV therefore means that the solutions collectively dominate a larger portion of the objective space. ENS achieves appreciable improvements over state-of-the-art baselines on bi-objective problems and ENS-Aug is also the only neural method in Table 1 to outperform WS-LKH on Tri-TSP.

Each individual ENS solver uses the standard POMO architecture. The longer inference time reported for ENS in some settings comes primarily from evaluating more solvers. For example, ENS retained 660 solvers in our original Bi-TSP50 run. To make the comparison fairer, we ran additional experiments on Bi-TSP50 using the strongest single-model and multi-model baselines, WE-CA and EMNH, respectively. We increased the number of preference queries for WE-CA from 101 to 701 using the same trained model, and increased the number of preference-specific models for EMNH from 101 to 701. The table below reports HV, inference time, and total offline computation time for all three methods on Bi-TSP50. The reported offline time includes anchor training and evolutionary extension for ENS, 200 training epochs using the original paper’s hyperparameters for WE-CA, and estimated meta-training plus measured fine-tuning for EMNH.

Bi-TSP50 with 64-fold augmentation

MethodQueriesModelsInitial trainingFine-tune / evolveOffline totalHV ↑Inference ↓
WE-CA10118.75 h—8.75 h0.64123.3 m
20118.75 h—8.75 h0.64185.23 m
30118.75 h—8.75 h0.64207.82 m
40118.75 h—8.75 h0.642110.41 m
50118.75 h—8.75 h0.642113.00 m
60118.75 h—8.75 h0.642215.60 m
70118.75 h—8.75 h0.642218.25 m
EMNH—1019.58 h7.55 m9.71 h0.64081.7 m
—2019.58 h10.41 m9.76 h0.64214.91 m
—3019.58 h13.42 m9.81 h0.64277.35 m
—4019.58 h16.81 m9.86 h0.64299.80 m
—5019.58 h20.80 m9.93 h0.643112.24 m
—6019.58 h23.15 m9.97 h0.643314.69 m
—7019.58 h26.43 m10.02 h0.643417.14 m
ENS—6605.00 h56.23 m5.94 h0.64357.5 m

Increasing WE-CA from 101 to 701 queries raises HV from 0.6412 to 0.6422, while inference time grows from 3.3 to 18.25 m. EMNH benefits more from additional models, reaching an HV of 0.6434 with 701 solvers, close to ENS. ENS has the lowest Inference time at 7.5min and also the lowest total offline time at 5.94 h, compared with 8.75 h for WE-CA and an estimated 9.71–10.02 h for EMNH. Evolutionary search accounts for about 56 m of ENS’s offline time and expands five trained anchors into an ensemble that achieves the highest HV in this comparison.

Q2. Variability across evolutionary runs

We repeated the evolutionary stage ten times on Bi-TSP50 with independent random seeds, using the same five trained anchors and the same 200 test instances. With augmentation, the mean HV was 0.6436, the sample standard deviation was 2.82×10−52.82 \times 10^{-5}, and the 95% Student's tt confidence interval had a half-width of 2.02×10−52.02 \times 10^{-5}. Without augmentation, the mean HV was 0.6399, the sample standard deviation was 6.40×10−56.40 \times 10^{-5}, and the 95% confidence interval had a half-width of 4.58×10−54.58 \times 10^{-5}. These intervals describe uncertainty in the mean HV across evolutionary runs with fixed anchors and test instances.

AuthorsOpenReview ↗

Response to Reviewer Sz9G (Part 2/2)

Q3. Ablation of evolutionary components

We ran additional ablations on Bi-TSP50 with the same five anchors, random seed, and budget of 100 generations for each variant. Without direction estimation removes the estimated directional offset from mutations. Without reseeding uses the anchors to initialize the population, then generates offspring from the existing parent pool. Independent selection within each anchor applies diversity-based survivor selection separately to each anchor’s population, so candidates from different anchors do not compete. Reseeding only generates all offspring by applying independent Gaussian perturbations to the anchors’ query projections. It retains joint survivor selection and the archive, while disabling directional guidance and the use of archived elites as parents.

Evolutionary variantHV without augmentation ↑HV with augmentation ↑
Full ENS0.64000.6437
Without direction estimation0.63910.6435
Without reseeding0.63900.6436
Independent selection within each anchor0.63880.6427
Reseeding only0.63410.6407

Full ENS achieves the highest HV with and without augmentation. Replacing joint selection with independent selection within each anchor lowers HV in both settings, as does removing direction estimation or reseeding individually. With direction estimation disabled in both variants, retaining elite selection raises augmented HV from 0.6407 to 0.6435, isolating its contribution under that configuration.

Thank you for clarifying the offline training costs and reporting the results from repeated evolutionary runs. On Bi-TSP50, ENS reaches nearly the same hypervolume (HV) as the 701-model EMNH configuration in less than half the inference time. Compared with the fastest reported setting for each baseline, ENS improves HV by 0.36% over WE-CA and 0.42% over EMNH, but takes 127% and 341% more inference time, respectively. At comparable inference times (7.5 versus 7.35 minutes), ENS improves HV by only 0.12% over the 301-model EMNH configuration. These results clarify the trade-off on Bi-TSP50, but it remains unclear whether the gains justify the additional inference time in other settings where ENS is slower than the baselines. The new ablations help clarify the contribution of the evolutionary components. Compared with the reseeding-only variant, which still uses evolutionary search and joint selection, full ENS improves HV by 0.93% without augmentation and 0.47% with augmentation. These gains seem modest, and without runtime comparisons for the ablation variants, it is difficult to assess whether the benefits of the individual components justify their added complexity. Thus, I am inclined to maintain my current score.

Reviewer_PLDEOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ENS, a two-state method for MOCO. ENS trains a small set of preference-specific neural solvers and then evolves their decoder query projections to generate diverse solutions. Experiments on several benchmark problems show the superiority of ENS.

Critical strengths
  • Centering the interpolation process around a knee point is a reasonable design choice.
  • The paper provides thorough ablation studies.
Critical weaknesses
  • Among the many possible interpolation and extension strategies, the paper does not sufficiently explain why optimizing the decoder’s query projection is particularly effective.
  • The experimental protocol for computing the evaluation metrics is insufficiently specified.
Questions for authors

In the evolutionary extension stage, the paper proposes mutating the decoder’s query projection. Although the empirical results shows the superiority of ENS, the rationale for this choice remains insufficient. Given the range of possible interpolation and extension strategies, could the authors provide a stronger analytical justification, or a theoretical guarantee under suitable assumptions, for why optimizing this projection is particularly effective?

Appendix B specifies the HV definition and reference points, but the solution-set construction remains unclear. For each method, how many models, preference vectors, and trajectories are evaluated, and which resulting solutions are retained for HV computation?

Since approximation-set cardinality can affect HV, an additional comparison using matched solution-set sizes would help separate solution quality from the benefit of returning more solutions: https://ieeexplore.ieee.org/document/7743917

Comparisons under matched computational budgets would also help assess the trade-off between solution quality and computational cost.

Finally, how are the number and distribution of evaluation preference vectors determined for the single-model methods? These choices can affect Pareto-front coverage and HV and should be explicitly reported.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.