acceptodds
Under review as a conference paper at ICLR 2027

No Sigmoid Required: Exact Feasibility for Constrained EHVI

Abstract

Constrained multi-objective optimization is essential for engineering design, requiring sample-efficient batch evaluation strategies due to costly simulations. Although Bayesian optimization using CEHVI is highly effective, estimating its probability of feasibility (PoF) relies on a sigmoid relaxation, which introduces bias and a tunable parameter. Na\"ive marginalization can eliminate the sigmoid, but we find that it can lead to overestimation of the acquisition function for when running points are spatially concentrated. This paper theoretically analyzes this error's source and derives an unbiased, low-variance Rao-Blackwellized (RB) estimator that replaces the candidate feasibility indicator with its exact conditional expectation. Using a synthetic problem that reproduces a concentrated scenario, we first verify that the RB approach suppresses the overestimation in the infeasible region. Furthermore, our approach resolves the sigmoid-induced PoF overestimation, improving sample efficiency near the feasibility boundary. Experiments on real-world benchmark problems demonstrate that our method yields stable performance, eliminating the parameter sensitivity of the sigmoid relaxation.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
AutoML, Hyperparameter Optimization, Bayesian Optimization, Gaussian Process, Black-Box Optimization, Multi-Objective Optimization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_SLqWOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies the estimation of constrained expected hypervolume improvement (qCEHVI) in batch multi-objective Bayesian optimisation. Existing implementations use a sigmoid relaxation to approximate feasibility indicators, which introduces bias and requires a temperature parameter. The authors show that directly replacing the sigmoid with the marginal probability of feasibility can also introduce bias when the batch size is greater than one. To address this issue, they propose a Rao-Blackwellized estimator based on conditional feasibility probabilities. The proposed estimator is shown to be unbiased, differentiable almost everywhere, and to have lower variance than the hard-indicator estimator. Experiments on synthetic and real-world constrained optimisation problems demonstrate its effectiveness and stability across different batch sizes.

Critical strengths
  1. The paper identifies a clear problem with using marginal PoF in batch constrained qEHVI. The authors explain why this can introduce bias and provide theoretical analysis to support their findings.
  2. The proposed method is simple and does not require tuning the sigmoid temperature parameter. The authors also prove that the estimator is unbiased, differentiable, and has lower variance than the hard-indicator estimator. The experiments show promising results across different batch sizes.
Critical weaknesses
  1. The experimental evaluation is limited to only five benchmark problems. It is unclear why these particular problems were selected. All of them have only two or three objectives, and the number of decision variables ranges from 2 to 10. The paper does not show whether the method works well on more challenging problems with higher dimensions and more objectives.
  2. The comparison with existing BO methods is limited. The paper does not include ParEGO or constrained ParEGO, which would be relevant baselines. Instead, it includes evolutionary algorithms such as NSGA-II and NSGA-III, which typically use relatively large populations and may not perform well under small evaluation budgets. Comparisons with more constrained BO methods would provide a better understanding of the proposed method's performance.
  3. The paper does not report the computational time of the proposed method. Although Section 3.4 states that RB and sigmoid have similar computational complexity, it would be useful to compare their actual running times, especially for larger batch sizes and more constraints.
Questions for authors
  1. Why were these five benchmark problems selected? Could the authors explain whether they cover the main challenges that the proposed method aims to address?
  2. How does the proposed method perform on problems with more decision variables and objectives? The current experiments only consider problems with 2 to 10 decision variables and 2 or 3 objectives.
  3. How does the proposed method compare with other constrained BO methods, such as constrained ParEGO? Would these methods provide more meaningful comparisons under small evaluation budgets?
  4. How does the actual running time of RB compare with the sigmoid approach? Does the difference become larger as the batch size or number of constraints increases?
  5. How does the proposed method perform when the constraints are correlated? The current experiments only consider independent GPs.
  6. How does the variance reduction in Theorem 2 affect the Monte Carlo estimation in practice? Could the authors compare the estimator variance and estimation error under different numbers of Monte Carlo samples?
  7. How does the marginal-PoF bias change with the distance or correlation between running points? Is this bias mainly a problem when running points are highly concentrated?
  8. Does the proposed method provide similar benefits when using joint batch optimisation instead of greedy batch selection?
3 comments

Thank you for your detailed reviews. We summarize the updates to the manuscript:

  • (Updated; Fig. 4) Benchmarked qqParEGO and qqHVKG.
  • (New; Fig. 9) Empirical variance reduction.
  • (New; Fig. 10) High-dimensional benchmarking.
  • (New; Fig. 11) Multi-output GP toy problem.
  • (New; Fig. 13) Per-AF call runtime benchmarking.
  • (New; Table 1) RB form per setup.
  • (New; Limitations) Discussion of missing evaluations.

Experiment Coverage

W2&Q3: BO Baselines

We added qParEGO and qHVKG [4] to Fig. 4.

Q6: Variance Reduction

Thm. 2 (iii) gives the per-sample variance gap E[HVI2p(1−p)]\mathbb{E}[\mathrm{HVI}^2 p(1-p)] (pp: conditional PoF). For iid draws, the gap is divided by SS (Cor. 11), so the advantage persists for large SS. In practice, RB benefits most when pp is far from 0 and 1 (feasibility is uncertain). While our QMC makes the theory approximate, new Fig. 9 shows RB's standard deviation at S=32S=32 is already below the hard indicator's at S=128S=128 on most slices.

W3&Q4: Runtime Report

We added per-call AF runtime report (Fig. 13). The runtime difference between sigmoid and RB is indistinguishable, consistent with Sec. 3.4. Marginal PoF is slightly faster because it uses an unconditional rather than conditional GP. Pipeline runtime also depends on AF optimization iteration counts (Fig. 6).

W1&Q1/2: Problem Selection

We clarified our problem selection rationale in the manuscript. Handling low feasibility is crucial, so our setup pairs one relatively easy problem (Disc Brake) with several strict ones. Multi-product Batch Plant (0.6% feasibility), and Process Synthesis (<10−4<10^{-4}%) represent strict real-world constraints. We made sure that mixed space and tri-objective problems are also covered. TNK highlights how marginal PoF bias degrades performance compared to RB. SRN is a strict synthetic function.

While high-dimensional and many-objective analyses are valuable, comprehensive studies fall outside our scope. Nevertheless, to address your concern, we added two high-dimensional problems. MW7 (d=15d=15) is taken from the BoTorch test function suite where MW7 & C2DTLZ2 are the only high-dimensional CMO options and MW7 has far lower feasible ratio. DASCMOP2 (difficulty 2) is taken from pymoo where it was the highest-dimensional choice and difficulty 2 is controlled for low feasible ratio. RB is competitive with sigmoid (Fig. 10). However, we believe that a more detailed analysis is required to develop better designs for high-dimensional problems, which we leave for future work.

Regarding the number of objectives, evaluating M=2M=2 or 33 is standard MOBO practice. Recent BoTorch literature [1-4] does not test constrained many-objective (M>3M>3) problems (though [3] tested unconstrained M=4M=4). Since our core focus is enhancing constraint handling, we believe the current scope of our evaluation is appropriate.

Q7: Marginal-PoF Bias

The bias -Cov[HVI, I] (Prop. 1) grows with the posterior correlation between the candidate's and running points' constraints. Prop. 14 shows bias vanishes without correlation. It is bounded by V[HVI]p(1−p)\sqrt{V[\mathrm{HVI}] p(1-p)} (marginal PoF: pp), maximizing near the feasibility boundary (Rem. 15). Concentration isn't required, as one correlated running point suffices. This makes all cross-covariances nonnegative, turning the bias into a one-sided overestimate (Prop. 19) pulling the batch toward the cluster.

Q5/8: Joint Batch Optimization & Correlated Constraints

Thm. 2 (appendix Thms. 9, 10, 17) and Prop. 1 (Thm. 12) remain valid under joint optimization and correlated constraints. Thus, exactness, variance-reduction, differentiability, and marginal-PoF bias results carry over. Only the closed-form univariate expression (Cor. 21, Rem. 25) is lost. Table 1 summarizes this, and limitations note the joint-batch case lacks empirical evaluation.

For correlated constraints, we added an experiment (Fig. 11) using TNK and modified TNK with duplicated constraints. Results show negatively correlated constraints degrade RB's performance with multi-output GP (aligning with Zhang et al. [5]), but duplicating a constraint reversed this. We note the limited scope in the text, but hope this resolves your question.

References

[1] Ament et al. Unexpected Improvements to Expected Improvement for Bayesian Optimization. NeurIPS, 2023

[2] Daulton et al. Differentiable Expected Hypervolume Improvement for Parallel Multi-Objective Bayesian Optimization. NeurIPS, 2020

[3] Daulton et al. Multi-Objective Bayesian Optimization over High-Dimensional Search Spaces. UAI, 2022

[4] Daulton et al. Hypervolume knowledge gradient: a lookahead approach for multi-objective Bayesian optimization with partial information. ICML, 2023

[5] Zhang et al. Dependence in constrained Bayesian optimization: When do we need it and how does it help? Optimization Letters, 2024.

Thank you for the additional experiments and detailed clarifications, which have addressed most of my concerns. I will maintain my positive score.

AuthorsOpenReview ↗

Thank You for Your Response and Positive Assessment

Thank you for your response and for your positive assessment. We're glad the additional experiments and clarifications addressed most of your concerns.

Reviewer_6oKKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies feasibility estimation in constrained EHVI for batch Bayesian optimization (qCEHVI). It analyzes the bias introduced by sigmoid relaxation and shows that naively multiplying expected improvement by the marginal probability of feasibility can overestimate the acquisition function for (Q>1) under certain correlation conditions. To address these issues, the authors propose a Rao–Blackwellized (RB) estimator and establish its unbiasedness, reduced variance relative to the hard-indicator estimator, and almost-everywhere differentiability. Experiments demonstrate competitive and stable performance without temperature tuning, while highlighting the sigmoid approach’s sensitivity to its temperature parameter.

Critical strengths

The paper is well motivated and provides a clear theoretical analysis of the limitations of marginal feasibility weighting and sigmoid relaxation.

The experiments complement the analysis by illustrating these failure modes and supporting the proposed estimator’s robustness.

Critical weaknesses

My main concern is the limited methodological novelty. The paper focuses on probability-of-feasibility estimation within the existing qCEHVI and sequential greedy selection framework. The unbiasedness and variance reduction follow directly from standard Rao–Blackwell arguments.

The presentation could more clearly. The paper switches between two sources of bias: the dependence ignored by marginal feasibility weighting and the approximation introduced by sigmoid relaxation. it's difficult to track which limitation each theoretical result or experiment addresses.

Minor error: line 355 claims all three methods are perfectly identical for Q=1. However, for a fixed, positive temperature, sigmoid relaxation generally remains biased even when (Q=1).

Questions for authors

Can you also include the comparison of wall clock time, will RB estimator requires longer running time compared with sigmoid or Margianl PoF?

In its experiments, the main performance evaluation relies on hypervolume. While complementary metrics, such as inverted generational distance and measures of solution diversity, would help assess the improvements.

4 comments
AuthorsOpenReview ↗

Clarifications on Critical Weaknesses & Corresponding Revisions

We sincerely thank you for the time and effort you have devoted to reviewing our work and for your constructive feedback.

W1 (novelty)

We agree that unbiasedness and variance reduction stem from standard Rao–Blackwellization (RB) arguments. These standard arguments themselves are not the methodological novelty we claim. Our contribution is to uncover a failure mode of the widely used sigmoid relaxation (Prop. 3; Fig. 3) and develop a principled remedy while preserving the batch dependence ignored by a naive replacement with marginal PoF (Prop. 1; Fig. 2). The novelty lies in this new understanding of the failure mechanisms and their simple, principled resolution, even though the mathematical tools are well established.

First, the sigmoid relaxation introduced by Daulton et al. (NeurIPS 2020, NeurIPS 2021) [1,2] is widely used in major Bayesian optimization frameworks such as BoTorch and Optuna [3,4] and real-world applications [5]. We show that it can favor candidates with large EHVI even when the GP confidently predicts them to be infeasible. Sigmoid relaxation assigns a positive feasibility weights even when the GP are conficent about infeasibility (Prop. 3; Fig. 3). This is particularly problematic for the important class of constrained optimization problems whose optima lie near feasibility boundaries [6-8]. It can cause repeated infeasible evaluations and prolonged hypervolume stagnation (Figs. 4–5). Our method resolves this failure without sacrificing differentiability.

Second, removing sigmoid relaxation is not straightforward: a naive replacement with marginal PoF introduces a different bias for Q>1Q>1, despite being exact at Q=1Q=1 under objective–constraint independence. Running-point fantasies induce dependence between the candidate's HVI and feasibility, which marginal PoF ignores. We characterize this bias and its sign (Props. 1 and 19) and demonstrate the resulting acquisition function (AF) overestimation near concentrated running points (Fig. 2). Our conditional PoF preserves this dependence.

Additionally, establishing sufficient conditions for the sign of marginal-PoF bias (Prop. 19, using Kahane's inequality) and proving pathwise differentiability (Theorem 2(ii); Theorem 17) require nontrivial analysis beyond standard RB identities.

Regarding sequential greedy batch selection, our theory does not require this restriction (Section 3.4; Table 1, Appendix A). We adopt this setting in the main text because it is a well-established standard in BO [1,4,9,10], as discussed in Section 2.2.

W2 (presentation issues)

We clarified the links between the two biases, theoretical results, and experiments. Changes are highlighted in blue.

Linking theory to the two biases (Sec. 3). The opening of Sec. 3 now explicitly links marginal-PoF bias to Prop. 1, sigmoid relaxation bias to Prop. 3, and RB's resolution of both to Thm. 2. We also revised Sec. 3.3 to specify the sigmoid failure mode it analyzes.

Making each experiment’s purpose explicit (Sec. 4). To clarify which limitation each experiment tests, we added explicit experimental questions as paragraph headings in Sec. 4.2. To further improve presentation, we also clarified our overall experimental aim at the beginning of Sec. 4 (to justify replacing sigmoid with RB), and added a paragraph at the end of Sec. 4.1 explaining what each benchmark problem is intended to test.

Regarding Minor error: line 355, the opening of Sec. 4 now clarifies that all three implementations use analytic LogCEHVI at Q=1Q=1 for fairness.

AuthorsOpenReview ↗

Actions Made for Reviewer 6oKK's Questions & References

Q1 (runtime)

We added mean wall-clock time per candidate proposal (Fig. 8), cumulative proposal time (Fig. 12), and per-AF-call runtime (Fig. 13; the latter two in Appendix C), covering all five benchmarks (L=2L=2 – 1010) and Q∈4,8,16,32Q\in\\{4,8,16,32\\}. Proposal times include both GP fitting and AF optimization; per-call runtimes measure the time required for a single evaluation of the AF value and its gradient.

RB and sigmoid have similar per-call costs across these settings (Fig. 13), consistent with Section 3.4. At the proposal level, smaller τ\tau generally increases sigmoid's proposal time (Figs. 8 and 12). This is because their gradients are unstable and requires more AF calls for convergence (Fig. 6), whereas RB's stable gradients avoid this increase. The shorter sigmoid proposal times at large τ\tau on TNK and SRN stem from stagnation caused by repeated infeasible proposals (Figs. 4,5).

Marginal PoF is somewhat faster per AF call because it skips feasibility conditioning on running-point fantasies, ignoring HVI–feasibility dependence and potentially biasing the AF estimate for Q>1Q>1 (Prop. 1; Fig. 2).

Q2 (IGD and Pareto-front coverage)

We added IGD histories (Fig. 14; Appendix C.4) and Pareto-front visualizations (Figs. 15, 16; Appendix C.5) to show solution diversity.

The IGD histories cover all five benchmarks and Q∈4,8,16,32Q\in\\{4,8,16,32\\}. RB often achieves lower (better) IGD than sigmoid and marginal PoF, especially at larger QQ, while sigmoid again stagnates on TNK and SRN.

The Pareto-front plots show the four bi-objective problems at the paper's evaluation budgets using seed 42. RB collects more nondominated feasible points in many settings, with visibly denser front coverage on TNK and SRN; sigmoid instead makes more infeasible evaluations concentrated at similar objective-space locations. At larger QQ, RB obtains more nondominated feasible points and makes fewer infeasible evaluations than marginal PoF in several settings.

These results suggest that, by avoiding the sigmoid bias toward confidently infeasible points (Fig. 3; sigmoid issue) and accounting for running-point dependence that can otherwise favor redundant candidates (Fig. 2; marginal PoF issue), RB can improve Pareto-front coverage more efficiently within the same evaluation budget.

References

[1] Samuel Daulton, Maximilian Balandat, and Eytan Bakshy. Differentiable expected hypervolume improvement for parallel multi-objective Bayesian optimization. NeurIPS, 2020.

[2] Samuel Daulton, Maximilian Balandat, and Eytan Bakshy. Parallel Bayesian optimization of multiple noisy objectives with expected hypervolume improvement. NeurIPS, 2021.

[3] BoTorch: Official documentation of acquisition functions. https://botorch.readthedocs.io/en/stable/acquisition.html

[4] Optuna v5.0.0: Implementation of GP acquisition functions (including qLogCEHVI). https://github.com/optuna/optuna/blob/v5.0.0/optuna/_gp/acqf.py

[5] Evolution-guided Bayesian optimization for constrained multi-objective optimization in self-driving labs. npj Computational Materials, 2024. https://www.nature.com/articles/s41524-024-01274-x

[6] Victor Picheny. A stepwise uncertainty reduction approach to constrained global optimization. AISTATS, 2014.

[7] Zhi-Zhong Liu, Bing-Chuan Wang, and Ke Tang. Handling constrained multiobjective optimization problems via bidirectional coevolution. IEEE Transactions on Cybernetics, 52, 2022.

[8] Yunsheng Tian, Ane Zuniga, Xinwei Zhang, Johannes P. Dürholt, Payel Das, Jie Chen, Wojciech Matusik, and Mina Konaković Luković. Boundary exploration for Bayesian optimization with unknown physical constraints. ICML, 2024.

[9] James Wilson, Frank Hutter, and Marc Deisenroth. Maximizing acquisition functions for Bayesian optimization. NeurIPS, 2018.

[10] Takashi Wada and Hideitsu Hino. Bayesian optimization for multi-objective optimization and multi-point search. arXiv:1905.02370, 2019.

Thanks for your response and additional experiment results. These have addressed my concerns of novelty, and I have updated my score accordingly.

AuthorsOpenReview ↗

Thank You for Your Response and Positive Assessment

Thank you very much for your thoughtful response and positive reassessment of our paper. We sincerely appreciate your constructive feedback and are glad that our response and additional results have addressed your concerns.

Reviewer_5prxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a Rao–Blackwellized (RB) estimator for constrained q-Expected Hypervolume Improvement (qCEHVI), replacing the biased sigmoid relaxation with conditional probabilities of feasibility to achieve unbiased, low-variance, and differentiable estimation. Theoretical analysis and experiments demonstrate that the proposed method addresses the bias caused by batch-induced dependencies and improves optimization stability without requiring temperature tuning.

Critical strengths
  • The paper provides a clear theoretical analysis of the bias introduced by marginal feasibility estimation in batch optimization.

  • The proposed estimator eliminates the sigmoid temperature parameter while maintaining differentiability. The experiments cover multiple benchmark problems and batch sizes.

Critical weaknesses
  • The proposed method mainly applies the Rao–Blackwellization technique and standard Gaussian conditioning to the existing qCEHVI framework. The theoretical properties, including unbiasedness and variance reduction, largely follow from standard conditional expectation results. Although the analysis of batch-induced dependence is useful, the overall contribution appears incremental, and I am not fully convinced that the methodological novelty is sufficient for ICLR.

  • The paper aims to replace the sigmoid relaxation, but the practical benefits of this replacement seems to be limited. The proposed estimator does not consistently outperform the sigmoid baselines, raising questions about its practical advantages. Although the authors investigate temperature sensitivity, the results do not demonstrate significant improvements over well-tuned sigmoid alternatives under comparable computational budgets. A more direct comparison of optimization performance and computational cost would help establish the practical value of the proposed approach. Evaluations on more challenging real-world applications involving expensive black-box simulations would further demonstrate its practical significance and scalability.

Questions for authors

Please refer to the weaknesses.

3 comments

We sincerely thank the reviewer for the time and effort you have devoted to reviewing our work and for your constructive feedback.

W1 (novelty)

We agree that Rao–Blackwellization and Gaussian conditioning are classical techniques. However, the novelty of our work lies not in these techniques themselves. We identify an overlooked failure mode in the widely used sigmoid approach. The seemingly natural fix of replacing sigmoid relaxation with marginal PoF is itself biased for Q > 1 (Prop. 1; Fig. 2). We believe that uncovering these overlooked problems, providing new methodological understanding of their causes, and resolving them with a simple solution constitute meaningful novelty, even when the mathematical tools are established.

First, we identify a concrete failure mode of the sigmoid relaxation introduced by Daulton et al. (NeurIPS 2020, NeurIPS 2021) [1,2] and widely used in practice, including in major Bayesian optimization frameworks such as BoTorch and Optuna [3,4] and in real-world applications [5].

Specifically, we identify a failure mode in which the sigmoid relaxation can favor candidates with large EHVI even when the GP confidently predicts them to be infeasible. This occurs because sigmoid relaxation assigns a positive feasibility weight even as the GP becomes confident about infeasibility, potentially leading to maximizers of the acquisition function (AF) in infeasible regions (Prop. 3; Fig. 3). This failure mode is particularly detrimental to the important class of constrained optimization problems whose optima lie near feasibility boundaries [6,7,8]. Our experiments further demonstrate the corresponding degradation in HVI (Fig. 4), including repeated infeasible evaluations (Fig. 5). Our method resolves the issue without sacrificing differentiability.

Second, eliminating the sigmoid relaxation is not as straightforward as simply replacing it with the marginal probability of feasibility (marginal PoF). While this approach is exact for Q=1Q=1 under objective–constraint independence, it introduces bias for Q>1Q>1 induced by running-point fantasy samples. We characterize this bias theoretically (Props. 1 and 19) and demonstrate how it leads to AF overestimation near spatially concentrated running points (Fig. 2). Our proposed estimator resolves this issue by preserving the dependency through conditional PoF.

Also, beyond standard RB properties (Theorem 2 (i), (iii)), establishing sufficient conditions for the sign of the marginal-PoF bias characterized in Proposition 1 (Prop. 19, using Kahane's inequality) and proving the pathwise differentiability stated in Theorem 2(ii) (Theorem 17) require nontrivial theoretical analysis and does not immediately follow from from standard conditional expectation results.

W2 (practical benefits and computational cost)

Regarding the practical benefits, our method effectively avoids sigmoid's prolonged hypervolume stagnation on TNK and SRN (Fig. 4; failure discussed in W1). Fig. 5 supports this explanation, showing that sigmoid's feasible ratio declines toward zero, whereas ours recovers.

Furthermore, our method also eliminates problem-dependent temperature tuning. Large τ\tau degrades performance on TNK and Disc Brake, whereas small τ\tau degrades performance on Multi-product Batch Plant and performs significantly worse than RB (Fig. 7; Wilcoxon signed-rank test, p<0.05p < 0.05). Importantly, this problem-dependent τ\tau is unknown in advance. With sufficient budget, tuning τ\tau may improve performance on a particular problem, but requires multiple additional BO runs, and an unsuitable choice can substantially degrade performance. RB provides stable performance competitive with sigmoid at a well-tuned τ\tau, without this tuning cost.

Moreover, under matched AF evaluation budgets, RB achieves substantially higher performance than sigmoid across all tested τ\tau (Fig. 6, left). Although small τ\tau reduces the smoothing bias, it requires more AF evaluations to converge (or gets stuck near a poor initialization when τ=10−6\tau = 10^{-6}) (Fig. 6, right). Our method avoids this trade-off between accuracy and AF optimization cost, thanks to the exactness and stable gradients.

To further quantify the computational cost, we added mean wall-clock proposal times (Fig. 8) and cumulative proposal times (Fig. 12, Appendix C) across all five benchmarks and Q∈4,8,16,32Q\in\\{4,8,16,32\\}. Both include GP fitting and AF optimization. Fig. 13 (Appendix C) isolates the time for a single AF value-and-gradient evaluation.

Smaller τ\tau generally increases sigmoid's proposal time through additional AF calls, whereas RB's stable gradients avoid this increase (Figs. 8 and 12). Their per-call times are similar (Fig. 13), consistent with Section 3.4. Sigmoid's shorter proposal times at large τ\tau on TNK and SRN stem from repeated infeasible proposals and hypervolume stagnation (Figs. 4–5), so shorter runtimes alone in these settings do not indicate better optimization performance.

AuthorsOpenReview ↗

References to the last response

References

[1] Samuel Daulton, Maximilian Balandat, and Eytan Bakshy. Differentiable expected hypervolume improvement for parallel multi-objective Bayesian optimization. NeurIPS, 2020

[2] Samuel Daulton, Maximilian Balandat, and Eytan Bakshy. Parallel Bayesian optimization of multiple noisy objectives with expected hypervolume improvement. NeurIPS, 2021

[3] BoTorch: Official documentation of acquisition functions. https://botorch.readthedocs.io/en/stable/acquisition.html

[4] Optuna v5.0.0: Implementation of GP acquisition functions (including qLogCEHVI). https://github.com/optuna/optuna/blob/v5.0.0/optuna/_gp/acqf.py

[5] Evolution-guided Bayesian optimization for constrained multi-objective optimization in self-driving labs. npj Computational Materials, 2024. https://www.nature.com/articles/s41524-024-01274-x

[6] Victor Picheny. A stepwise uncertainty reduction approach to constrained global optimization. AISTATS, 2014.

[7] Zhi-Zhong Liu, Bing-Chuan Wang, and Ke Tang. Handling constrained multiobjective optimization problems via bidirectional coevolution. IEEE Transactions on Cybernetics, 52, 2022.

[8] Yunsheng Tian, Ane Zuniga, Xinwei Zhang, Johannes P. Dürholt, Payel Das, Jie Chen, Wojciech Matusik, and Mina Konaković Luković. Boundary exploration for Bayesian optimization with unknown physical constraints. ICML, 2024.

AuthorsOpenReview ↗

Update: Additional results (IGD and Pareto-front coverage)

Here are additional results on solution diversity quantification to further verify the practical benefits of our method.

To provide further evidence of RB's practical benefits beyond HV, we added IGD histories (Fig. 14; Appendix C.4) and Pareto-front visualizations (Figs. 15, 16; Appendix C.5). Across all five benchmarks and Q∈4,8,16,32Q\in\\{4,8,16,32\\}, RB often achieves lower (better) IGD than the sigmoid baseline (τ=0.01\tau=0.01 as in Fig. 4) and marginal PoF, especially at larger QQ. Sigmoid again stagnates on TNK and SRN, whereas RB continues to improve its approximation of the reference front.

The Pareto-front plots compare the four bi-objective problems using seed 42 at the paper's evaluation budgets. RB collects more nondominated feasible points in many settings, with visibly denser front coverage on TNK and SRN, while sigmoid makes more infeasible evaluations concentrated at similar objective-space locations. The advantage over marginal PoF also becomes more pronounced at larger QQ. These results are consistent with the HV gains: by avoiding the sigmoid bias toward confidently infeasible points (Fig. 3; sigmoid issue) and accounting for running-point dependence that can otherwise favor redundant candidates (Fig. 2; marginal PoF issue), RB can improve Pareto-front coverage more efficiently within the same evaluation budget.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Manuscript Update Summary

We greatly appreciate all the reviewers' time and effort in reading our paper. Based on their insightful feedback, we have revised our manuscript accordingly. All our text updates are highlighted in blue for convenience.

Here is a high-level summary of what we've done:

  • (Updated; Sec. 3, top) Correspondence between issues and theory.
  • (Updated; Sec. 4, top) Clarification of how we implemented all three qqCEHVI variants at Q=1Q=1.
  • (Updated; Sec. 4.1) Problem choice rationale.
  • (Updated; Fig. 4) Benchmarking of qqParEGO and qqHVKG.
  • (Updated; Sec. 4.2) Paragraph headers to make each experiment purpose clearer.
  • (New; Fig. 8) Per-BO-iteration runtime benchmarking.
  • (New; Fig. 9) Empirical variance reduction.
  • (New; Table 1) RB form per setup.
  • (New; App. C) Additional experiments.
    • (New; Fig. 10) High-dimensional benchmarking.
    • (New; Fig. 11) Multi-output GP toy problem.
    • (New; Fig. 12) Cumulative runtime benchmarking.
    • (New; Fig. 13) Per-AF call runtime benchmarking.
    • (New; Fig. 14) IGD history plot for 20 settings used in Fig. 4.
    • (New; Figs. 15, 16) Pareto front visualization for bi-objective problems for Q=4,8,16,32Q=4,8,16,32.
  • (New; App. D) Discussion of missing evaluations.

We will keep this list updated as we make further revisions during the rebuttal period.

Updates:

  • 13 Nov. 2026: Add Figs. 14–16.

Dear Reviewers and Area Chair,

With less than 24 hours remaining in the discussion period, we would greatly appreciate any feedback on our revisions. In particular,

  • 5prx: We identified failure mechanisms in the current de facto (the sigmoid approach) and its seemingly natural fix (marginal PoF), so our contribution isn't centered on methodology but rather on concept. We believe that the conceptual contribution is sufficiently novel to benefit the community to better understand the downsides of the sigmoid approach. The "well-tuned τ\tau" is an oracle. The best τ\tau differs across problems (Fig. 7) and needs an outer loop of extra BO runs, which expensive evaluations rule out. RB matches it without tuning. Do these address your novelty concern?

To AC, we briefly summarize the current status:

  • 5prx (conf. 2, rating 2): We clarified the practical benefits: competitive performance with no τ\tau tuning (the best τ\tau differs across problems and an unsuitable τ\tau can substantially degrade the sigmoid's performance; Fig. 7). Additionally, we clarified that our novelty lies not in using the established math tools (RB and related theory) but in identification of the failure mechanisms (for both the existing method and its seemingly natural fix) and their simple, principled resolution.
  • 6oKK (conf. 3, rating 3): We fixed presentation issues in Sec. 3,4 and added IGD experiments. The reviewer confirmed that its novelty concern was resolved. (Update: rating 2 -> 3)
  • SLqW (conf. 4, rating 3): The reviewer raised no novelty concern, and we addressed most experiments (ParEGO, high-dimension, variance etc). The reviewer confirmed that it is happy about our response.
  • All: We added runtime benchmarks (Figs. 8, 12, 13).

We sincerely appreciate the reviewers' time and consideration, and would welcome any remaining questions or concerns.

AuthorsOpenReview ↗

Final summary

Dear the Area Chair,

We briefly summarize the final outcome of the rebuttal:

  • SLqW (conf. 4, rating 3; positive assessment maintained): The reviewer raised no novelty concern. We addressed most of their experimental concerns (ParEGO, high-dimensional benchmarks, variance, runtime, etc.). The reviewer confirmed that our additions addressed most of their concerns and maintained their positive score.
  • 6oKK (conf. 3, rating 3; novelty concern resolved): We clarified presentation issues in Secs. 3–4 and added IGD experiments and runtime benchmarks. The reviewer explicitly confirmed that their novelty concern was resolved. (Update: rating 2 → 3)
  • 5prx (conf. 2, rating 2; no follow-up response): We clarified the practical benefits: competitive performance without τ\tau tuning (the best τ\tau differs across problems, and an unsuitable τ\tau can substantially degrade sigmoid performance; Fig. 7), as well as improved performance under matched AF evaluation budgets (Fig. 6). We also clarified that our novelty lies not in the established mathematical tools (RB and related theory), but in identifying the failure mechanisms of both the current de facto and its seemingly natural fix, and resolving them in a simple, principled way.

We greatly appreciate your time for assessing our paper and rebuttal.

Warm regards, Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.