acceptodds
Under review as a conference paper at ICLR 2027

Beyond the Reference Trap: Adaptive Reverse Tchebycheff Scalarization for Pareto Set Learning in Expensive Multi-Objective Optimization

Abstract

Pareto set learning (PSL) learns a continuous mapping from preference vectors to Pareto-optimal solutions, providing an appealing approach to expensive multi-objective optimization under limited evaluation budgets. Most existing scalarization-based PSL methods typically estimate reference points from the evaluated archive. We show that this simple design choice can induce an initialization-dependent reference geometry that restricts Pareto-front exploration: when the archive covers only a local portion of the front, the resulting scalarization can bias the learned solutions toward the observed region and make unexplored trade-offs difficult to reach. We refer to this failure as the reference trap. To address it, we propose Adaptive Reference Geometry for PSL, instantiated through Adaptive Reverse Tchebycheff (AdaRevTche). Instead treating the reference as a fixed archive statistic, AdaRevTche jointly learns the Pareto-set mapping and a bounded worse-side reference, allowing the scalarization geometry to adapt as optimization proceeds. We establish an exact result showing how the reference point restricts the reachable region in two-objective Tchebycheff scalarization, and analyze the corresponding geometric behavior for problems with more objectives. Results on 12 real-world and synthetic problems demonstrate the effectiveness of our AdaRevTche. Notably, compared with the SoTA PSL baseline, CDM-PSL, AdaRevTche reduces the hypervolume difference on RE41 by 51.8%.

Primary area:
optimization
Keywords:
Multi-objective optimization, Pareto set learning, Bayesian optimization, Tchebycheff scalarization

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 3 reviews, above 27% of the others.
Reviewer_3sQvOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that Tchebycheff-based Pareto set learning (PSL) suffers from a "reference trap": when the ideal point is estimated from a local archive, global scalarized optima stay inside an archive-induced box. The proposed AdaRevTche replaces the ideal point with a worse-side (nadir) reference in a reverse Tchebycheff (max-min margin) scalarization. The reference is shifted outward by a bounded, softmax-parameterized offset that is learned jointly with the PSL model and reused for greedy HVI batch selection.

Critical strengths
  • Reference-point sensitivity in Tchebycheff scalarization-based PSL is a real, under-discussed practical issue, and the paper states it precisely.
  • The ablations are useful: Tche vs. RevTche vs. AdaRevTche, the range BB, and adapting the reference in PSL training vs. HVI selection. They come with paired confidence intervals.
Critical weaknesses
  1. Missing key related work; limited novelty of the scalarization. Without ρ\rho, Eq. (6) is a monotone transform of the hypervolume scalarization sλ(y)=min⁡i(ri−yiλi)ms_\lambda(y) = \min_i \left( \frac{r_i - y_i}{\lambda_i} \right)^m (Golovin & Zhang, 2020). That scalarization is already used for PSL in "Hypervolume Maximization: A Geometric View of Pareto Set Learning" (Zhang et al., 2023). What remains new is the bounded adaptive offset, which is a modest contribution.
  2. Missing simple baselines. The decisive missing baselines are (a) Tchebycheff with a utopian/ideal offset, i.e., z=z^∗−ϵz = \hat{z}^* - \epsilon, and (b) reverse/HV scalarization with a fixed offset, e.g., r=r^+δr = \hat{r} + \delta or the standard r=1.1×r^r = 1.1 \times \hat{r} HV reference rule, where r^\hat{r} is the archive nadir. Without these, the benefit of learning the reference is not established.
  3. Overclaimed results. Contribution 3 says AdaRevTche "consistently achieves lower hypervolume differences … across all considered tasks," and the conclusion makes a similar claim. Table 2 and Section 4.2 contradict this: it is not best on F2, RE21, or convex polar, and RE31, RE35, and RE37 are statistically tied.

References

Daniel Golovin and Qiuyi Zhang. Random hypervolume scalarizations for provable multi-objective black box optimization. In International Conference on Machine Learning (ICML), 2020.

Xiaoyuan Zhang, Xi Lin, Bo Xue, Yifan Chen, and Qingfu Zhang. Hypervolume maximization: A geometric view of Pareto set learning. In Advances in Neural Information Processing Systems (NeurIPS), 2023.

Questions for authors

Please see weaknesses.

2 comments

We appreciate your time and consideration in reviewing our work. We provide our responses to your comments below.


W1. Missing key related work; limited novelty of the scalarization. Without , Eq. (6) is a monotone transform of the hypervolume scalarization (Golovin & Zhang, 2020). That scalarization is already used for PSL in "Hypervolume Maximization: A Geometric View of Pareto Set Learning" (Zhang et al., 2023). What remains new is the bounded adaptive offset, which is a modest contribution.

Response. Thank you for pointing out these relevant works. We have added both references to the “Scalarization methods” paragraph in Section 2 and acknowledged the connection to prior hypervolume scalarizations.

Comparing to those two paper, our novelty lies in identifying the reference trap and developing adaptive reference geometry for PSL, instead to claim the scalarization itself as novel. We theoretically characterize how archive-estimated references can restrict Pareto-front exploration, even with exact objectives and globally solved subproblems. We then jointly learn a worse-side reference and the PSL map to alleviate this restriction and enable exploration beyond the observed front.


W2. Missing simple baselines. The decisive missing baselines are (a) Tchebycheff with a utopian/ideal offset, i.e., , and (b) reverse/HV scalarization with a fixed offset, e.g., or the standard HV reference rule, where is the archive nadir. Without these, the benefit of learning the reference is not established.

Response. We have added fixed-margin comparisons in Section 4.3 (Figure 6), testing δ∈{0.01,0.1,0.2}\delta\in\{0.01,0.1,0.2\} for Tchebycheff training, RevTche training, and HVI selection. AdaRevTche achieves lower final mean HV differences than all tested fixed-margin variants on RE41, RE42, RE61, and ZDT1, while fixed margins perform slightly better on Convex Polar. The benefit of a fixed margin varies across problems, and larger margins do not necessarily improve performance, supporting adaptive reference adjustment.

Table A. Final log HVD, ln⁡(HV∗−HVˉ)\ln(HV^*-\bar{HV}), where HVˉ\bar {HV} is the mean final HV over ten runs; lower is better. Each fixed-offset variant reports its best tested δ\delta in parentheses; bold marks the best per task.

ProblemTche + δ\deltaRevTche + δ\deltaHVI + δ\deltaAdaRevTche(Ours)
RE414.59 (0.2)4.71 (0.2)4.71 (0.2)4.24
RE4224.94 (0.1)24.97 (0.1)24.73 (0.2)24.56
RE6169.00 (0.1)68.95 (0.01)68.95 (0.2)68.77
ZDT1−1.03 (0.01)−2.37 (0.1)−2.06 (0.01)−3.16
Convex Polar−5.53 (0.1)−5.50 (0.2)−5.54 (0.2)−5.48

W3. Overclaimed results. Contribution 3 says AdaRevTche "consistently achieves lower hypervolume differences … across all considered tasks," and the conclusion makes a similar claim. Table 2 and Section 4.2 contradict this: it is not best on F2, RE21, or convex polar, and RE31, RE35, and RE37 are statistically tied.

Response. We have revised the wording to reflect the task-level results. AdaRevTche wins or ties in 90 of 93 Holm-adjusted comparisons and significantly outperforms every available baseline on RE41, RE42, and RE61, the more challenging, higher-dimensional engineering tasks (Appendix D, Table 3 and Figure 11).


We appreciate the opportunity to clarify our approach and hope the additional evidence answers your questions.

AuthorsOpenReview ↗

Following Up on Our Responses — Reviewer 3sQv

Dear Reviewer 3sQv,

Thank you again for your helpful comments. We have provided detailed responses, clarified our contribution and its relationship to prior work, and added the requested fixed-margin comparisons.

We look forward to hearing your thoughts and would be glad to address any outstanding concerns.

Best regards, The Authors

Reviewer_w9yiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses scalarization-based Pareto set learning (PSL) for expensive multi-objective optimization. It argues that estimating the Tchebycheff ideal point from the evaluated archive induces a "reference trap". For two objectives, every global Tchebycheff minimizer lies within the box bounded by the coordinate-wise maxima of the archive (Prop. 3.1/A.1), and a conditional version is given for m > 2 (Prop. A.2). The proposed AdaRevTche anchors a reverse Tchebycheff scalarization at the archive nadir and adds a bounded, softmax-parameterized outward offset. This offset is learned jointly with the set model and is also used as the reference point for greedy HVI batch selection. Experiments cover eight RE engineering problems, three synthetic benchmarks (ZDT1, F2, VLMOP3) and a newly constructed Convex Polar problem. Each setting uses 110 evaluations per run and 10 runs, and the eight baselines include PSL-MOBO, DA-PSL, CDM-PSL, DGEMO and qLogEHVI.

Critical strengths

S1. The method is simple and adds little cost: m extra parameters per BO round. Yet it gives the best final HV on the three problems with four or more objectives, RE41, RE42 and RE61 (Table 2). On RE41 it reaches 1.06e3 ± 6.04, compared with 990 ± 15.8 for CDM-PSL; on RE42 it reaches 8.66e11 ± 5.78e9, compared with 8.15e11 ± 2.98e10. It also improves on PSL-MOBO in mean on all twelve problems. These results are the main reason I see practical value in the approach.

S2. The paper makes a useful point for PSL: the reference point restricts which part of the front is reachable, not only which trade-off is targeted. Prop. A.1 is correct and its proof is short, and Remark 2 is candid that the results hold in objective space only. The F/A ablation in Fig. 5b separates the effect of the reference in PSL training from its effect in HVI selection, and it reports paired confidence intervals. This makes the proposed mechanism testable.

Critical weaknesses

W1. The reference trap, and the gains of AdaRevTche, may largely reflect the use of zero-margin references. The baseline needed to rule this out is missing.

Theory. Prop. A.1 relies on ẑ_i being exactly the archive minimum, so that a^(1) has a zero first term. Once the ideal point is offset to a utopian point ẑ − δ, the containment no longer holds. This agrees with the classical result that weighted Tchebycheff with a utopian reference can reach every Pareto-optimal point (Miettinen, 1999). The same zero-margin effect appears in the two other references used in the pipeline. RevTche anchored at the archive nadir (B = 0) is confined to {y ≤ r̂}, up to the O(ρ) term. HVI with its reference at the nadir gives no credit to candidates beyond it (cf. Ishibuchi et al., Evol. Comput. 2018). The expansion therefore appears to come from the outward margin, not from reversing the scalarization.

Method. Sec. 3.3 shows only that ∂S_rev/∂r > 0, i.e., that S_rev is monotone in r. This explains why gradient ascent moves r outward, but not why the learned offset is a good one. Its final value is determined by B, β, the learning rate and the number of steps I. Moreover, by Eqs. (7) and (9), R(a) is minimized exactly when softmax(a) is uniform, and there all displacements vanish. The regularizer by itself therefore favors no offset (see Q5).

Experiments. The in-house Tchebycheff variant reaches HV 48.63 on RE21 (App. C), compared with 57.5 for the official PSL-MOBO (Table 2). The ablations in Fig. 5a and App. C are therefore made against a weak baseline. AdaTche, which learns the ideal point by gradient descent on T, is not an informative comparator either. T is decreasing in z, so descent moves z away from the utopian region and toward the solutions, and App. C itself shows that this contracts the sublevel box. No fixed-margin baseline is reported.

Experiment needed. The missing comparison is with fixed-margin baselines on RE41, RE42, RE61 and Convex Polar:

  • Tchebycheff with ẑ − δ·range;
  • RevTche with r̂ + δ·range;
  • an HVI reference at the nadir + δ·range;

each for δ ∈ {0.01, 0.1, 0.2}. If AdaRevTche does not clearly outperform these, the contribution reduces to adding a margin. This also bears on novelty, since inverted scalarizations and worse-side references are already established in decomposition-based EMO (Sato 2014; Jiang & Yang 2016, both cited).

W2. The empirical claims are stronger than Table 2 supports.

Contribution 3 states that AdaRevTche "consistently achieves lower HV differences than PSL-MOBO and non-PSL baselines across all considered tasks". Table 2 shows otherwise:

  • DGEMO has the best mean on F2 and RE35;
  • DA-PSL has the best mean on RE21;
  • CDM-PSL has the best mean on RE37 and Convex Polar.

The abstract highlights the best case (RE41, −51.8%), while AdaRevTche is not best on 5 of the 12 problems. On Convex Polar, which was built to isolate the trap, the gap between PSL-MOBO and AdaRevTche is small (0.891 vs. 0.896, with H* = 0.900), and CDM-PSL is best. The main comparison also lacks significance tests; with 10 runs, many differences are within one standard deviation (F2, RE21, RE31, RE35, RE37). Rank-sum tests with Holm correction and a win/tie/loss summary would help, and the abstract, contributions and conclusion should be revised accordingly.

Questions for authors

Q1. Are the objectives normalized before Eqs. (6)–(9) and before HVI selection? If not, B = 1 is negligible relative to the objective scales of RE31, RE42 and RE61 (10³–10⁷). It is large, however, for the second objective of RE21 (scale ≈ 0.04), even though all hyperparameters are fixed across problems.

Q2. Does the official PSL-MOBO implementation set the ideal point to the exact archive minimum or use an offset? Which reference point does it use for HVI? The answer would help explain the gap between your Tchebycheff variant (48.63) and PSL-MOBO (57.5) on RE21.

Q3. In the diagnostic of Fig. 1 and App. B.2, candidates are drawn from the dense exact front and selected by greedy HVI. At which step is Tchebycheff scalarization applied, and what HVI reference point is used?

Q4. The abstract refers to CDM-PSL as the state-of-the-art PSL baseline. However, Sec. 4.1 states that only PSL-MOBO and SVH-PSL learn Eq. (3), and Sec. 4.2 describes DA-PSL as non-Tchebycheff. Which baselines use an archive-estimated Tchebycheff ideal point and are therefore subject to the trap?

Q5. Sec. 3.2 describes R(a) as encouraging "a diverse allocation of the reference displacement". Under the centered clip in Eq. (7), however, the minimizer of R (uniform softmax(a)) gives zero displacement on every objective. For m = 2, at most one objective can be displaced in a given round. Is this the intended behavior? What displacements are typically learned, e.g., on RE41 and RE61?

3 comments

We appreciate your thoughtful review and constructive suggestions. Below, we explain how we addressed your concerns.


W1.1 The reference trap, and the gains of AdaRevTche, may largely reflect the use of zero-margin references. The baseline needed to rule this out is missing.

Reponse. The performance gains of our method are primarily from adaptive reference learning, rather than zero-margin references. The original baselines already incorporate reference adjustments: Tche includes the official implementation’s 0.9z∗0.9z^* ideal-point offset, RevTche uses 1.1n^1.1\hat n as its training reference, and HVI uses the coordinatewise maximum over evaluated points and candidate predictions. All these operations are performed in standardized objective space.

Although a fixed outward margin can partially alleviate the reference trap, its effectiveness varies across tasks. Different problems exhibit different Pareto-front geometries and exploration requirements, making it difficult for a fixed margin to consistently provide an appropriate reference. In contrast, AdaRevTche adaptively adjusts the reference across objectives according to the evolving optimization process, enabling more flexible exploration of underexplored Pareto regions.

To verify this, we conducted additional fixed-offset experiments with δ∈{0.01,0.1,0.2}\delta\in\{0.01,0.1,0.2\}. These offsets are applied in addition to the original reference : Tchebycheff with 0.9z^∗−δ0.9\hat z^* - \delta, RevTche with 1.1n^+δ1.1\hat n + \delta, HVI reference at the n^+δ\hat n + \delta.

Table A. Final log HVD, ln⁡(HV∗−HVˉ)\ln(HV^*-\bar{HV}), where HVˉ\bar {HV} is the mean final HV over ten runs; lower is better. Each fixed-offset variant reports its best tested δ\delta in parentheses; bold marks the best per task.

ProblemTche + δ\deltaRevTche + δ\deltaHVI + δ\deltaAdaRevTche(Ours)
RE414.59 (0.2)4.71 (0.2)4.71 (0.2)4.24
RE4224.94 (0.1)24.97 (0.1)24.73 (0.2)24.56
RE6169.00 (0.1)68.95 (0.01)68.95 (0.2)68.77
ZDT1−1.03 (0.01)−2.37 (0.1)−2.06 (0.01)−3.16
Convex Polar−5.53 (0.1)−5.50 (0.2)−5.54 (0.2)−5.48

We have included these results in Figure 6 of Section 4.3 in the revised manuscript.


W1.2 Sec. 3.3 shows only that ∂Srev/∂r>0\partial S_{rev}/\partial r > 0, i.e., that SrevS_{rev} is monotone in rr. This explains why gradient ascent moves rr outward, but not why the learned offset is a good one.

Reponse. Learning the offset provides a signal to increase the reference margin of the current bottleneck, allowing improvements in other objectives to influence optimization. Ignoring the augmentation term, Srev=min⁡i(ri−yi)/λiS_{\mathrm{rev}}=\min_i(r_i-y_i)/\lambda_i, and the reference gradient acts only on the unique bottleneck kk. Increasing rkr_k raises its margin; once it exceeds the second-smallest margin, the bottleneck switches to jj. For a feasible trade-off direction that worsens kk while improving jj, with vk>0v_k>0 and vj<0v_j<0, the directional derivative of SrevS_{\mathrm{rev}} then changes from −vk/λk<0-v_k/\lambda_k<0 to −vj/λj>0-v_j/\lambda_j>0. Thus, the learned offset can make previously suppressed directions favorable, changing the preference over trade-off directions.


W2. The empirical claims are stronger than Table 2 supports. Rank-sum tests with Holm correction and a win/tie/loss summary would help, and the abstract, contributions and conclusion should be revised accordingly.

Response. Thank you for your valuable suggestion. We have carefully revised the experimental descriptions to more accurately reflect the task-level results and avoid overstating the empirical advantages of AdaRevTche.

To further assess the reliability of the observed improvements, we have conducted additional statistical significance analyses and reported the results in Appendix D of the revised manuscript (Table 3 and Figure 11). Specifically, based on Holm-adjusted tests of the final per-run log HVD, AdaRevTche achieves 66 statistically significant wins and 24 statistically indistinguishable comparisons, accounting for 90 of the 93 pairwise comparisons. Notably, on the more challenging, higher-dimensional engineering tasks RE41, RE42, and RE61, AdaRevTche significantly outperforms all seven available baselines on each task.

Q1. Are objectives normalized before reference adaptation and HVI? How should B=1B=1 be interpreted?

Response. Yes. We apply objective normalization consistently across all compared methods. Specifically, at each BO round, each objective is centered by the mean and scaled by the standard deviation of the evaluated archive. The GP surrogate, PSL training objective in Eqs. (6)–(9), and HVI-based candidate selection all operate in this normalized objective space.

Accordingly, (B=1) defines the adaptation range in normalized objective units, rather than in the original objective scales. This allows us to use a consistent adaptation parameter across problems with different objective magnitudes. We have make it more clear in Section 4.1 of the revised manuscript.


Q2. Does the official PSL-MOBO implementation set the ideal point to the exact archive minimum or use an offset? Which reference point does it use for HVI? The answer would help explain the gap between your Tchebycheff variant (48.63) and PSL-MOBO (57.5) on RE21.

Response. Official PSL-MOBO uses an ideal-point offset, and our fig3, fig5a follows its code setting, correspond to HV=57.48HV=57.48. The appendix removes the offset to examine the reference trap under the empirical-ideal assumption, corresponding to HV=48.63HV=48.63. The paper specifies z⋆−0.1∣z⋆∣z^\star-0.1|z^\star|, while the code uses a 0.10.1 offset in standardized space. HVI selection uses 1.1n1.1n where the coordinatewise maximum over evaluated points is nn.


Q3. In the diagnostic of Fig. 1 and App. B.2, candidates are drawn from the dense exact front and selected by greedy HVI. At which step is Tchebycheff scalarization applied, and what HVI reference point is used?

Response. In the candidate-generation step, we minimize Tchebycheff over the dense exact front for each sampled preference to form the candidate pool. The reference is the current archive ideal with the outward correction z^−0.1∣z^∣\hat z-0.1|\hat z|.

Greedy HVI then selects up to five candidates using the fixed oracle reference (1.1,1.1)(1.1,1.1) in the original objective space. Using the exact front isolates reference-induced reachability from surrogate error. These steps are clarified in Appendix B.2 of the revised manuscript.


Q4. The abstract refers to CDM-PSL as the state-of-the-art PSL baseline. However, Sec. 4.1 states that only PSL-MOBO and SVH-PSL learn Eq. (3), and Sec. 4.2 describes DA-PSL as non-Tchebycheff. Which baselines use an archive-estimated Tchebycheff ideal point and are therefore subject to the trap?

Response. PSL-MOBO and SVH-PSL use archive-derived ideal references in Tchebycheff-based training; qNParEGO also uses Chebyshev scalarization in its acquisition function. DA-PSL and CDM-PSL generate solutions to cover the Pareto front without learning an explicit preference-to-solution mapping.


Q5. Sec. 3.2 describes R(a) as encouraging "a diverse allocation of the reference displacement". Under the centered clip in Eq. (7), however, the minimizer of R (uniform softmax(a)) gives zero displacement on every objective. For m = 2, at most one objective can be displaced in a given round. Is this the intended behavior? What displacements are typically learned, e.g., on RE41 and RE61?

Response. Yes, this behavior is intentional. The regularizer encourages a diverse allocation of reference adjustments by discouraging excessive allocation to any single objective. It favors uniform allocation, which corresponds to zero displacement. For m=2m=2, the centered parameterization and clipping allow at most one objective to receive a positive displacement in each round.

To illustrate the learned reference allocation, we recorded the mean signed offsets at round 20 with B=1B=1 in diagnostic reruns:

  • RE41: (+0.062,+0.123,+0.051,−0.236)(+0.062,+0.123,+0.051,-0.236)
  • RE61: (−0.147,−0.082,+0.259,+0.134,−0.030,−0.133)(-0.147,-0.082,+0.259,+0.134,-0.030,-0.133)

These are pre-clipping signed offsets, negative values are clipped to zero under Eq. (7). We have included the corresponding trajectories in Appendix F.1 (Figures 12 and 13) in revised manuscript.


Your comments helped us identify where the paper needed clearer explanations. We hope the revisions address your concerns.

AuthorsOpenReview ↗

Looking Forward to Further Discussion — Reviewer w9yi

Dear Reviewer w9yi,

We sincerely appreciate your careful review and constructive suggestions. We have responded to your questions and added experiments and analyses to address the concerns you raised.

We would welcome your thoughts on the additional evidence and would be happy to discuss any remaining questions.

Best regards,

The Authors

Reviewer_46KYOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a method to find a representative covering of the Pareto set. One standard way of parametrizing the Pareto front is through the Tchebycheff scalarization. This parametrization depends on a hyperparameter: a reference point. Ideally, the reference point should dominate all feasible solutions. For example, it could be the vector of outcomes that would be achieved if it were possible to simultaneously be optimal on all objectives without the need for trade-offs. The issue, which this paper points out, is that we may not know the ideal point and need to use an estimate. If done poorly, the Tchebycheff scalarization with respect to a bad reference point will fail to parametrize the whole Pareto set.

This paper uses this to motivate their approach, which makes use of a reverse Tchebycheff scalarization. Intuitively, one way to think about the Tchebycheff scalarization is that it grows a weighted ℓ∞\ell^\infty-ball centered at the reference point, and each weighting correspond to the first feasible point that the sequence of balls hit. The reverse Tchebycheff is similar, except that it places the reference point on the other side of the Pareto front, and the chain of balls are given with respect to the sum of a weighted ℓ∞\ell^\infty and ℓ1\ell^1-norm.

The paper also introduces an algorithm that attempts to produce a set of feasible solutions that can approximate the Pareto set well. Given a decision vector xx, the learner can make a query to an oracle that returns f(x)=(f1(x),…,fm(x))f(x) = (f_1(x),\ldots, f_m(x)) the tuple of values. These calls are presumably expensive, so the goal is to minimize the number of function evaluations. They take a Bayesian optimization approach: the maintain a dataset D={(x,f(x)),…}D = \{(x,f(x)),\ldots\} of evaluated decisions. They use a Gaussian process to estimates a surrogate for ff. Using the reverse-Tchebycheff scalarization on this surrogate objectives, this helps inform them how to choose the next points to evaluate ff on. They validate this algorithm on some benchmarks.

Critical strengths

This paper studies a fundamental problem.

Critical weaknesses

A major issue for me is that the paper is rather difficult to read. I don't think the argument about the "reference trap" really justifies why you need the reverse Tchebycheff scalarization (nor do I understand why the reverse scalarization fundamentally overcomes an analogous reference trap issue, but on the other side). More specifically:

  1. It seems that the reference trap is mainly pointing out that having a bad reference point will eliminate certain points from being included in a learned Pareto set. But, if this is the issue, why would the solution not be more direct? For example, you could also try to learn how to minimize the objectives at the same time you're trying to learn how to minimize their trade-offs. It is also not clear why the reverse Tchebycheff would be helpful, except that there is this additional ℓ1\ell_1-term. Would that also help if added to the usual Tchebycheff scalarization (e.g. Section 2.4 of [1], or Section 3.4.5 of [2])?
  2. There are a lot of what I think are non-standard terms used without introduction. For example, line 166 says that D\mathcal{D} is the "evaluated archive". While this later becomes clear with the algorithm, it is undefined at this point in the paper. Your later claim after Proposition 3.1 that "an archive-induced empirical reference can confine scalarized optima to regions determined by the current archive, preventing subsequent evaluations from expanding toward uncovered trade-offs" (line 190). But this seems to presume some sort of algorithm (e.g. what are initial/current/subsequent archives?). But this algorithm is never introduce (perhaps it is implied by your own algorithm?)
  3. The language in Proposition 3.1 is perhaps too informal. What does it mean for a subgradient to "point strictly toward a box"?
  4. I could not really parse the exposition leading up to the algorithm. I think the point about Gaussian process needs more explanation. The way it currently reads, the algorithm seems a bit arbitrary. Overall, it would be nice if you could provide more justification for your design choices.

[1] Ehrgott (2006), A discussion of scalarization techniques for multiple objective integer programming. https://link.springer.com/article/10.1007/s10479-006-0074-z [2] Miettinen (1999). Nonlinear Multiobjective Optimization

Questions for authors

I would recommend the authors really work on exposition. While I understand roughly what your paper does, I don't see why you've made your design choices, nor how they are contributing to better performance. As I am not a practitioner, I am not sure how to think about the experiments you've done, and whether they really suggest that this method always provides a clear advantage. Perhaps it would be better to shed more light on when you expect this method would outperform other methods (and also when it might not).

3 comments

Thank you for taking the time to review our work. We will revise the manuscript to make the method easier to follow and explain the rationale for our design choices more clearly. We address each of your comments below.


W1.1. It seems that the reference trap is mainly pointing out that having a bad reference point will eliminate certain points from being included in a learned Pareto set. But, if this is the issue, why would the solution not be more direct? For example, you could also try to learn how to minimize the objectives at the same time you're trying to learn how to minimize their trade-offs.

Response. Direct reference search can be less reliable because optimizing a GP fitted to sparse data can amplify prediction errors. AdaRevTche learns bounded reference offsets using gradients from PSL training.

In our experiments, directly estimating the ideal or nadir reference through single-objective optimization performs worse than AdaRevTche on the tested problems. Tche + ideal search minimizes each objective’s GP lower confidence bound to estimate the ideal reference. RevTche + nadir search uses the resulting single-objective solutions to estimate the nadir reference. Results use 10 paired runs with 110 evaluations each. We have added the experiment in Appendix C of the revised manuscript.

Table A. Comparisons of final log HVD (ln⁡(HV∗−HVˉ)\ln(HV^*-\bar {HV}), where HVˉ\bar {HV} is the mean final HV over ten runs; lower is better).

ProblemPSL-MOBOTche + ideal searchRevTche + nadir searchAdaRevTche (Ours)
RE414.884.804.394.24
RE4225.0524.7625.0324.56
RE6169.0969.4569.0368.77

W1.2. It is also not clear why the reverse Tchebycheff would be helpful, except that there is this additional ℓ1\ell_1-term. Would that also help if added to the usual Tchebycheff scalarization?

Response. The ℓ1\ell_1-term is already used in PSL-MOBO. Jointly learning the reference with ordinary Tchebycheff can contract the reachable scope, so we use reverse Tchebycheff to enable outward reference adaptation (Section 3.3). In our experiments, both the PSL-MOBO baseline in Figure 3 and the Tchebycheff baseline in Figure 5(a) already included the augmentation term.


Q2. There are a lot of what I think are non-standard terms used without introduction. For example, line 166 says that D\mathcal{D} is the "evaluated archive". While this later becomes clear with the algorithm, it is undefined at this point in the paper. Your later claim after Proposition 3.1 that "an archive-induced empirical reference can confine scalarized optima to regions determined by the current archive, preventing subsequent evaluations from expanding toward uncovered trade-offs" (line 190). But this seems to presume some sort of algorithm (e.g. what are initial/current/subsequent archives?). But this algorithm is never introduce (perhaps it is implied by your own algorithm?)

Response. The evaluated archive is the dataset of all input–objective pairs (x,f(x))(x,f(x)) collected through actual objective evaluations. It includes every evaluated point, including dominated ones. Thus, D0\mathcal D_0 contains the observations from the initial design, Dt\mathcal D_t contains all observations available at the start of round tt, and Dt+1\mathcal D_{t+1} also includes the observations collected during that round.

We assume the standard MOBO loop, in which a surrogate fitted to the current archive guides the next evaluations, and the resulting observations are used to update the surrogate. In each round, the algorithm fits a surrogate to the current archive Dt\mathcal D_t and uses it to select new inputs. It then evaluates those inputs using the black-box objective functions and appends the resulting input–objective pairs to Dt\mathcal D_t to form the archive Dt+1\mathcal D_{t+1} for the next round.

Q3. What does it mean for a subgradient to "point strictly toward a box"?

Response. We mean that the negative subgradient strictly decreases the squared Euclidean distance to the box to first order. Let At(λ)\mathcal A_t(\lambda) be the preference-conditioned attraction box in Definition 1 of Appendix A and let ΠAt(λ)(y)\Pi_{\mathcal A_t(\lambda)}(y) be the Euclidean projection of yy onto it. For every yy outside the box and every q∈∂yTt(y;λ)q\in\partial_y T_t(y;\lambda), Appendix A gives

=-\big(T_t(y;\lambda)-\tau_t(\lambda)\big)<0.$$ The global scalarized minimizers are also inside the box. --- > **Q4.** I could not really parse the exposition leading up to the algorithm. I think the point about Gaussian process needs more explanation. The way it currently reads, the algorithm seems a bit arbitrary. Overall, it would be nice if you could provide more justification for your design choices. **Response. Gaussian processes are a standard surrogate choice in MOBO because they provide both predictions and uncertainty estimates, with a posterior mean and standard deviation that can each be differentiated with respect to the candidate inputs.** These derivatives give us the gradient of the lower confidence bound for training the Pareto set model. We also use the surrogate values to score candidates before evaluating the selected batch. > **Question for authors.** When should we expect AdaRevTche to help, and when might the gains be small? **Response. We expect AdaRevTche to help when archive-estimated references restrict Pareto-front coverage;** gains may be smaller when surrogate accuracy or the problem’s inherent optimization difficulty is the main bottleneck. A reference estimated from a limited archive may favor already explored regions and make other trade-offs difficult to reach. AdaRevTche jointly learns the reference and the PSL map, adjusting the reference outward to allow exploration of these regions. When inaccurate predictions or difficulty in finding good solutions is the dominant limitation, changing the reference alone may offer limited improvement. --- We hope these responses clarify the points you raised. Thank you again for your feedback.
AuthorsOpenReview ↗

Looking Forward to Your Response — Reviewer 46KY

Dear Reviewer 46KY,

Thank you again for your valuable feedback and suggestions on improving the presentation of our work. We have provided detailed responses to your concerns and clarified the rationale behind our design choices.

We would appreciate hearing whether these responses address your concerns and welcome any further questions or suggestions.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Responses

Dear ACs and Reviewers,

We extend our sincere gratitude for your valuable time and insightful feedback on our paper. Your comments have been instrumental in refining our work. In addition to our specific responses to each reviewer, we would like to 1) express our gratitude for your recognition of our work, and 2) emphasize the major modifications made in our revised manuscript.

  1. We are encouraged by your acknowledgment of the importance and strengths of our work.

    • “This paper studies a fundamental problem.” [Reviewer 46KY] “Reference-point sensitivity in Tchebycheff scalarization-based PSL is a real, under-discussed practical issue, and the paper states it precisely.” [Reviewer 3sQv]
    • “The paper makes a useful point for PSL: the reference point restricts which part of the front is reachable, not only which trade-off is targeted.” “Prop. A.1 is correct and its proof is short, and Remark 2 is candid that the results hold in objective space only.” [Reviewer w9yi]
    • “The method is simple and adds little cost: m extra parameters per BO round.” [Reviewer w9yi]
    • “Yet it gives the best final HV on the three problems with four or more objectives, RE41, RE42 and RE61 (Table 2).” “It also improves on PSL-MOBO in mean on all twelve problems. These results are the main reason I see practical value in the approach.” [Reviewer w9yi]
    • “The ablations are useful” and “They come with paired confidence intervals.” [Reviewer 3sQv] “The F/A ablation in Fig. 5b separates the effect of the reference in PSL training from its effect in HVI selection, and it reports paired confidence intervals. This makes the proposed mechanism testable.” [Reviewer w9yi]
  2. We summarize the main modifications in our revised paper (highlighted in blue).

    • We add fixed-margin comparisons for Tchebycheff training, RevTche training, and HVI selection in Section 4.3 (Figure 6). The new results compare adaptive reference learning with additional fixed offsets and demonstrate that the effect of a fixed margin varies across problems. [Reviewers w9yi, 3sQv]
    • We add direct ideal- and nadir-reference search experiments in Appendix C (Figure 10 and Table 2), evaluating single-objective surrogate optimization as an alternative way to estimate the reference. [Reviewer 46KY]
    • We add Holm-adjusted statistical comparisons in Appendix D (Figure 11), reporting 66 significant wins, 24 statistically indistinguishable comparisons, and 3 significant losses. We revise the performance claims to reflect the task-level results. [Reviewers w9yi, 3sQv]
    • We clarify objective standardization in Section 4.1, candidate generation and batch selection for the exact-front diagnostic in Appendix B.2, and the empirical-ideal reference setting of the RE21 diagnostic in Appendix C. [Reviewer w9yi]
    • We add historical reference-offset trajectories in Appendix F.1 to illustrate the allocation mechanism and regularization comparisons in Appendix F.2 to examine the effect of the regularizer. [Reviewer w9yi]
    • We add the suggested hypervolume scalarization references to the “Scalarization methods” paragraph in Section 2 and acknowledge their connection to reverse Tchebycheff. [Reviewer 3sQv]

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.