- 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
- Combining neural solvers with evolutionary search effectively improves Pareto-front coverage.
- Restricting evolution to the decoder query projection is a simple and reasonable design choice, supported by the ablation results.
- The experiments cover multiple MOCO problems and include strong baselines such as WE-CA, with particularly promising results on tri-objective TSP.
Critical weaknesses
- 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.
- 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.
- 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.
- 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
- What are the key advantages of ENS over prior neural evolutionary MOCO methods under comparable computational budgets?
- 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?
- What is the total computational cost of evolution? How are multiple solvers executed at inference, and are encoder computations shared?
- How much do direction-guided evolution and global non-dominated selection contribute compared with random perturbations and conventional population selection methods?
3 comments
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 . The provided derivation merely shows that perturbing (W_q) changes attention scores, which is also true for other projections such as . It does not sufficiently explain why 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.
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 Wq. 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 Wq. 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.