- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes an evolution-aware state representation for deep-RL-based dynamic algorithm configuration (DAC) of multi-objective evolutionary algorithms. Existing DAC state encoders (e.g., GS-MODAC) represent only the current population, so populations at the same search stage with different rates of change look alike. The authors address this by (i) fitting multivariate Gaussians to consecutive populations' objective vectors and computing the closed-form KL divergence D_KL(P_{t-1}||P_t) as a scalar summary of inter-generational change, (ii) concatenating this normalized scalar with normalized search progress t/T into a 2-D "evolution-aware" feature e_t, (iii) mapping e_t through an MLP (tanh) to a feature-wise weight vector w_t that element-wise reweights the GCN node embeddings of a Pareto-dominance population graph, and (iv) concatenating the pooled representation with e_t as the policy state. A PPO agent outputs continuous crossover/mutation probabilities for NSGA-II, trained with a normalized HV-improvement reward. Validation covers Bi/Tri/Penta-FJSP and Bi-CVRP (12 settings), cross-size generalization, zero-shot cross-problem transfer (DAFJS-SDST variants, CVRPLIB Set X), an ablation, a reweighting visualization, and a runtime analysis.
Critical strengths
- Well-motivated, lightweight mechanism. Capturing recent evolutionary dynamics with a 2-D feature rather than a full temporal graph is appealing and empirically adds almost no test-time overhead (Table 7: within ±6%). This is consequential because it keeps DAC practical on larger instances.
- Broad and systematic evaluation. Two problem families, varying objective counts, cross-size/cross-problem transfer, real CVRPLIB Set X instances, ablation, and a visualization of the reweighting process together make a fairly thorough empirical study.
Critical weaknesses
- Marginal and possibly non-significant empirical gains, overstated as "significant." Improvements over GS-MODAC are tiny: Bi-FJSP-25j5m 9.51e4 vs 9.54e4, Tri-FJSP-25j5m 2.15e7 vs 2.16e7, Penta-FJSP-25j5m 5.62e12 vs 5.63e12, Bi-CVRP-100 1.35e2 vs 1.36e2. The only clearly larger margin is Bi-CVRP-500 (2.35e2 vs 2.42e2, ~3%). Yet the paper repeatedly claims "significant" improvements. Crucially, statistical testing is performed only against GS-MODAC (the underlined entries), not against irace or NSGA-II, and OURS's standard deviation is frequently higher. This directly undermines the central claim of superiority.
- The KL component is shown to contribute little, and the closest SOTA baseline is missing. In the ablation (Table 5), "OURS (w/o KL guidance)" nearly matches the full model (e.g., Bi-CVRP-200 1.61e2 vs 1.62e2; Bi-CVRP-500 2.41e2 vs 2.42e2), whereas "w/o feature reweighting" drops below GS-MODAC. This suggests the gain comes mainly from the generic reweighting mechanism, not the proposed evolution-aware/KL signal, and no significance test is given for the KL contribution. Moreover, STN-MODAC—the most closely related recent method explicitly modeling multi-generation dynamics—is discussed but never compared against, which is a serious omission for a novelty claim built on contrast with it.
Questions for authors
- Are the improvements over GS-MODAC/irace/NSGA-II statistically significant? Please report the Wilcoxon test results for all baselines, not only GS-MODAC, and clarify whether the sub-1% differences are significant.
- Given that "w/o KL guidance" is nearly identical to the full model, what is the demonstrable marginal contribution of the KL signal in isolation (with significance)? Does the reweighting mechanism alone explain most of the gain?
- Why is STN-MODAC not included as a baseline? How does the proposed method compare to it under identical budgets?
- What justifies modeling population objective vectors as Gaussians, and how sensitive are results to the covariance regularization ε=1e-6 and to the directional choice of KL?
- Why do MADAC and irace sometimes underperform fixed-parameter NSGA-II, and does this indicate tuning/training issues that affect fairness?
6 comments
Marginal Contribution of KL Guidance and Additional Ablation Evidence [W2 (part) / Q2]
We thank the reviewer for encouraging a clearer assessment of the respective contributions of KL guidance and feature-wise reweighting.
W2 (part): Removing KL guidance produces only small performance reductions, whereas removing feature-wise reweighting causes larger degradation, suggesting that the improvement may come primarily from feature reweighting.
Q2: What is the demonstrable marginal contribution of KL guidance, and does feature-wise reweighting alone explain most of the improvement?
We acknowledge that feature-wise reweighting contributes substantially to the performance improvements, particularly on Bi-CVRP, where removing KL guidance results in relatively small reductions in HV. However, this observation does not necessarily imply that the KL signal is redundant.
In our framework, feature-wise reweighting adaptively adjusts the contributions of latent node-embedding dimensions, while KL divergence supplies complementary information about inter-generational population changes. Their integration allows recent evolutionary dynamics to influence the construction of the population representation.
To further investigate their contributions beyond Bi-CVRP, we extended the ablation experiments to Penta-FJSP, covering three problem scales (5j5m, 10j5m, and 25j5m), as reported in Table 6 of Appendix D.3. Table 6: Ablation Study on Penta-FJSP
| Method | 5j5m Mean | 5j5m Max | 5j5m Std | 10j5m Mean | 10j5m Max | 10j5m Std | 25j5m Mean | 25j5m Max | 25j5m Std |
|---|---|---|---|---|---|---|---|---|---|
| GS-MODAC | |||||||||
| OURS (w/o feature-wise reweight) | |||||||||
| OURS (w/o KL guidance) | |||||||||
| OURS (FULL) |
The full model achieves the highest or tied-highest mean and maximum HV across these settings. Notably, on Penta-FJSP-5j5m, removing KL guidance reduces mean HV from to , whereas removing feature-wise reweighting results in . This indicates that the relative contributions of the two components can vary across optimization settings. Furthermore, the Wilcoxon signed-rank tests associated with Tables 5 and 6 indicate statistically significant improvements of the full model over GS-MODAC and both ablation variants in the evaluated Bi-CVRP and Penta-FJSP settings (p < 0.05).
To further examine whether Gaussian KL provides useful information beyond simpler temporal descriptors, we additionally compared it with mean displacement and Wasserstein distance on Bi-CVRP-200. Gaussian KL achieves the highest mean HV (162.38) and maximum HV (178.94), compared with mean displacement (160.15 and 176.28) and Wasserstein distance (161.71 and 177.99).
Table: Comparison of Alternative Temporal Metrics on Bi-CVRP-200
| Temporal Metric | Mean HV | Max HV | Std |
|---|---|---|---|
| Mean displacement | 160.15 | 176.28 | 9.53 |
| Wasserstein distance | 161.71 | 177.99 | 9.39 |
| Gaussian KL | 162.38 | 178.94 | 9.81 |
These results support the interpretation that KL guidance and feature-wise reweighting provide complementary benefits, although feature-wise reweighting may account for a larger share of the improvement in some settings.
Direct Comparison with STN-MODAC [W2 (part) / Q3]
We thank the reviewer for pointing out the need for a direct comparison with STN-MODAC under comparable experimental settings.
W2 (part): The absence of STN-MODAC, a closely related method that models multi-generational evolutionary dynamics, weakens the novelty and empirical comparison.
Q3: Why was STN-MODAC not included, and how does the proposed method compare with it under comparable computational budgets?
We agree that the absence of a direct STN-MODAC comparison was an important limitation of the original experimental evaluation.To address this concern, we added STN-MODAC as a baseline across nine FJSP settings, covering two, three, and five objectives at three different problem scales.
The overall comparisons are summarized in Table 1. Our method achieves the highest or tied-highest mean HV across these settings. Furthermore, Table 7 reports statistically significant improvements over STN-MODAC in all nine settings according to paired Wilcoxon signed-rank tests (p < 0.05). (Due to the character limit, Table 1 is provided in the next comment. Table 7 is presented in the first comment.)
Statistical Significance and Interpretation of Performance Improvements [W1 / Q1]
We thank the reviewer for emphasizing rigorous statistical testing and careful interpretation of modest performance improvements.
W1: Several improvements over GS-MODAC are numerically small and potentially non-significant. Statistical comparisons with other baselines are incomplete, and OURS sometimes exhibits higher standard deviations.
Q1: Are the improvements over GS-MODAC, irace, and NSGA-II statistically significant, particularly when the numerical differences are below 1%?
We acknowledge that several improvements over GS-MODAC are numerically modest, and we agree that the magnitude of an improvement should be distinguished from its statistical significance. Our original presentation could have more clearly differentiated these two aspects.
To provide stronger statistical evidence, we extended the two-sided paired Wilcoxon signed-rank tests to include NSGA-II and STN-MODAC, in addition to GS-MODAC. The results are reported in Table 7 of Appendix D.3.
Table 7: Wilcoxon Signed-Rank Test Results for Mean HV
Specifically, OURS achieves statistically significant improvements over GS-MODAC in 8 of the 12 settings, over NSGA-II in all 12 settings, and over STN-MODAC in all 9 evaluated FJSP settings (p < 0.05).
Notably, the four settings highlighted by the reviewer—Bi-FJSP-25j5m, Tri-FJSP-25j5m, Penta-FJSP-25j5m, and Bi-CVRP-100—yield p-values of 0.02367, 0.03911, 3.897×10−18, and 0.003832 against GS-MODAC, respectively. Although the numerical improvements are small in these cases, they are statistically significant under the reported tests. We nevertheless acknowledge that statistical significance does not necessarily imply a substantial practical improvement.
Regarding irace and MADAC, independently reproducing their results under precisely matched experimental conditions requires detailed implementation and hyperparameter settings, some of which could not be fully verified from the available descriptions. To maintain consistency with the established benchmark comparisons and avoid introducing discrepancies from potentially different configurations, we adopted the results reported in the GS-MODAC study. However, since the corresponding instance-level results are unavailable, statistical significance test cannot be reliably conducted. We acknowledge this limitation and do not claim statistically significant superiority over these two baselines. We also acknowledge that OURS does not consistently achieve lower standard deviations than GS‑MODAC. Therefore, our results should be interpreted as improvements in solution quality across several evaluated settings, rather than universal improvements in both performance and stochastic stability.