acceptodds
Under review as a conference paper at ICLR 2027

Converge Then Diversify: Decoupling Convergence and Diversity in Multi-Objective Bayesian Optimisation

Abstract

Multi-objective Bayesian optimisation (MOBO) is a sample-efficient approach for optimising expensive black-box functions with multiple objectives. In MOBO, the goal is to adequately approximate the Pareto front; that is, to obtain a high-quality solution set with 1) good convergence (closeness to the Pareto front) and 2) good diversity (spread across the Pareto front). Existing MOBO methods typically aim to accomplish these two tasks simultaneously, i.e., driving the search towards the Pareto front while maintaining a diverse set of nondominated solutions, such that the solutions, ideally, can gradually approach the entire front. When sufficient search budgets are available, this approach is effective. However, considering both convergence and diversity throughout the search is not easy and requires careful design. Under very tight budgets, there may not be enough solutions generated to be able to simultaneously approach the entire Pareto front. To address this issue, this paper proposes a converge-then-diversify (CTD) approach that decouples convergence and diversity into two stages. In the first stage, CTD focuses on convergence, aiming to quickly drive the search toward a single point on the Pareto front. In the second stage, CTD focuses on diversity, aiming to spread solutions across the front. We present two simple instantiations of CTD by using widely adopted acquisition functions in the area. Experimental results show that, across all 446 pairwise comparisons, CTD statistically outperforms state-of-the-art methods in 72.9% of the cases, performs equivalently in 21.1%, and is statistically worse in only 6.1%, with the advantage being particularly evident in settings with very tight evaluation budgets or in high-dimensional problems.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Bayesian Optimisation, Multi-objective Optimisation, Gaussian Process

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_yNsgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper present a new methods for MOBO problems. with finite function evaluations. The proposed searching method is flexible so that it can be used with different AC functions. Many problems show the effectiveness of the proposed method.

Critical strengths

Multi-objective optmization is an important problem.

Critical weaknesses
  1. The font size in the tables are too small, making the content difficult to read.

  2. The experimental evaluation is limited in scope, relying solely on long-established benchmarks suites such as ZDT and DTLZ. Given the advances in multi-objective optimization over the past two decades, the authors should include more recent and challenging benchmark problems to provide stronger evidence of the proposed method’s effectiveness.

  3. No real world problems.

  4. Too many words in Page 9, experiments are not so strong.

Questions for authors
  1. How is the reference points are selected in hv testing?

  2. How can you involve more real-word tesing problems?

  3. how is the performance on m=4,5,6,7? m=2/3 are just very easy to understand. too many existing works.

  4. What is the usage of Lem 1. It looks like that you have not built more theorems, so it is uncecssay to include lems.

5 comments
AuthorsOpenReview ↗

Response [1/3]

Q1: How are the reference points selected in the HV testing?

For the HV testing, we follow a commonly used setting in multi-objective optimisation [1]. Specifically, we collect all solutions obtained by all methods across all independent runs and set the reference point by extending the nadir point by 10% of the range between the ideal and nadir points in each objective.

We have clarified this in the "Performance Indicator" subsection of Section 5 in the revised manuscript.

Q2&W3: No real-world problems.

We would like to respectfully clarify that our experiments included six real-world optimisation problems: four-bar truss design, pressure vessel design, hatch cover design, vehicle safety design, car side-impact design, and laser plasma acceleration. These problems are widely used in the MOBO literature and cover different application domains [2-7]. For example, laser plasma acceleration [8] involves optimising three objectives, such as electron beam energy and charge, and represents a practical application of multi-objective optimisation in accelerator physics. We appreciate the reviewer's suggestion and would also be happy to consider including additional real-world problems to further strengthen the empirical evaluation.

AuthorsOpenReview ↗

Response [2/3]

Q3: Evaluation of CTD on problems with more objectives (m=4,5,6,7m=4,5,6,7).

CTD is expected to be competitive on many-objective problems. This is because CTD only considers convergence in the first stage, hence approaching the Pareto front substantially faster in a higher-dimensional space, whereas competing methods consider a number of directions simultaneously towards the Pareto front, resulting in slow convergence within a limited budget.

To confirm this, we have included the experiments on DTLZ problems with m=4,5,6,7m=4,5,6,7. Table A below presents the results. As shown, CTD achieves statistically better or equivalent HV than all competing methods on at least 26 out of 28 problems. We have now incorporated this comparison into the paper (Tables 3-6, page 26).

Table A: HV Results on DTLZ with m=4,5,6,7m=4,5,6,7.

MethodDTLZ1(4)DTLZ2(4)DTLZ3(4)DTLZ4(4)DTLZ5(4)DTLZ6(4)DTLZ7(4)
Sobol2.88e+05~1.47e+00+0.00e+00+5.43e-01+3.90e+00+1.91e+01+0.00e+00+
ParEGO1.06e+06~1.26e+00+9.06e+08~5.06e-01+4.56e+00+1.29e+02~3.23e-01~
TS-TCH1.74e+05~1.23e+00+0.00e+00+4.24e-01+4.19e+00+2.98e+01+0.00e+00+
PSL2.34e+06~1.35e+00+1.58e+09~5.30e-01+4.70e+00+1.33e+02~2.48e-01~
JES4.05e+05~1.13e+00+0.00e+00+3.40e-01+3.79e+00+2.56e+01+0.00e+00+
EHVI2.30e+06~1.18e+00+1.68e+09-3.19e-01+4.29e+00+1.33e+02~1.48e-01~
MORBO5.24e+05~1.35e+00+0.00e+00+3.85e-01+4.26e+00+3.10e+01+0.00e+00+
MOBO-OSD2.08e+06~9.63e-01+1.41e+09~5.85e-01+4.43e+00+1.34e+02~1.96e-01~
qPOTS3.09e+05~1.30e+00+0.00e+00+4.11e-01+4.07e+00+2.89e+01+0.00e+00+
CTD1.50e+062.51e+001.04e+091.65e+004.79e+001.34e+022.50e-01
MethodDTLZ1(5)DTLZ2(5)DTLZ3(5)DTLZ4(5)DTLZ5(5)DTLZ6(5)DTLZ7(5)
Sobol3.23e+07~3.98e-01+1.91e+10+5.30e-01+1.75e+01+3.82e+02+0.00e+00+
ParEGO1.33e+07~2.46e-01+2.18e+11+4.07e-01+2.02e+01~1.19e+03~3.80e-01+
TS-TCH3.00e+06+2.45e-01+0.00e+00+4.11e-01+1.86e+01+4.52e+02+0.00e+00+
PSL9.36e+07~3.68e-01+5.99e+11~5.20e-01+1.96e+01~1.19e+03~4.85e-01+
JES7.84e+06~2.46e-01+0.00e+00+3.04e-01+1.78e+01+3.97e+02+0.00e+00+
EHVI8.01e+07~3.26e-01+4.84e+11~3.47e-01+1.88e+01+1.20e+03~3.71e-01+
MORBO5.76e+06~3.99e-01+1.56e+10+4.74e-01+1.92e+01+4.35e+02+0.00e+00+
MOBO-OSD9.17e+07~4.65e-01+6.01e+11~4.20e-01+2.04e+01~1.18e+03~4.26e-01~
qPOTS3.09e+07~4.02e-01+0.00e+00+5.17e-01+1.89e+01+4.52e+02+0.00e+00+
CTD2.66e+071.30e+004.46e+111.36e+002.05e+011.19e+035.56e-01
MethodDTLZ1(6)DTLZ2(6)DTLZ3(6)DTLZ4(6)DTLZ5(6)DTLZ6(6)DTLZ7(6)
Sobol6.74e+09~6.16e+01+1.08e+13~4.32e+01+7.36e+00+4.16e+03+0.00e+00+
ParEGO5.67e+09~6.38e+01+8.99e+14~4.32e+01+7.88e+00+1.07e+04+5.68e-01~
TS-TCH6.25e+08~6.46e+01+3.88e+12~4.43e+01+7.79e+00+4.99e+03+0.00e+00+
PSL1.32e+10~6.03e+01+1.87e+15-4.12e+01+7.18e+00+1.07e+04+0.00e+00+
JES6.73e+09~6.13e+01+2.14e+13~4.29e+01+7.20e+00+4.12e+03+0.00e+00+
EHVI2.00e+10~6.14e+01+1.66e+15~4.18e+01+7.97e+00+1.08e+04+2.96e-01~
MORBO3.48e+09~6.47e+01+0.00e+00+4.44e+01+7.74e+00+5.29e+03+0.00e+00+
MOBO-OSD1.61e+10~6.02e+01+1.80e+15~4.39e+01+8.19e+00+1.09e+04+3.28e-01~
qPOTS1.45e+08~6.49e+01+3.95e+12~4.45e+01+7.88e+00+4.87e+03+0.00e+00+
CTD8.65e+096.94e+018.53e+144.80e+018.64e+001.09e+043.28e-01
MethodDTLZ1(7)DTLZ2(7)DTLZ3(7)DTLZ4(7)DTLZ5(7)DTLZ6(7)DTLZ7(7)
Sobol2.27e+13~3.14e+02~2.48e+20+2.11e+02+1.24e+01+3.86e+04+0.00e+00+
ParEGO9.99e+12~3.32e+02~3.75e+20~2.18e+02~1.35e+01+8.69e+04~2.91e-01+
TS-TCH1.43e+13~3.26e+02~2.73e+20+2.20e+02~1.35e+01+4.62e+04+0.00e+00+
PSL5.40e+13-3.00e+02~3.71e+20~2.03e+02+1.21e+01+8.48e+04+0.00e+00+
JES3.69e+12~3.14e+02~2.52e+20+2.10e+02+1.25e+01+4.14e+04+0.00e+00+
EHVI3.21e+13-2.98e+02~4.11e+20~2.03e+02+1.30e+01+8.61e+04~2.85e-01+
MORBO6.87e+12~3.28e+02~2.83e+20+2.20e+02~1.34e+01+4.70e+04+0.00e+00+
MOBO-OSD3.21e+13~3.00e+02~3.96e+20~2.03e+02+1.29e+01+8.60e+04~3.38e-01~
qPOTS1.43e+13~3.25e+02~2.74e+20+2.20e+02~1.34e+01+4.76e+04+0.00e+00+
CTD5.93e+123.33e+023.56e+202.25e+021.41e+018.70e+044.55e-01
AuthorsOpenReview ↗

Response [3/3]

Q4: What is the usage of Lemma 1?

A lemma is a mathematical result that can be used to support a theorem or provide theoretical motivation for a method [9]. In our paper, Lemma 1 is adopted from [10] to provide theoretical motivation for the EI-based switching criterion.

W1: The font size in the tables is too small.

To improve readability, we have reduced the standard deviations from two decimal places to one and tightened the spacing between the means and standard deviations.

W2: More recent and challenging benchmarks should be included.

In this work, we considered DTLZ benchmark problems as they are widely used in the MOBO literature [2-7]. That said, taking the reviewer's suggestion, we have now included the bi-objective BBOB benchmark suite [11]. BBOB is designed to capture common challenges in real-world optimisation, including separability, conditioning, and multi-modality [12], and its functions have been shown to exhibit diverse landscape characteristics representative of real-world problems [13-15].

The results are reported in Table 2 of the manuscript (page 25). As can be seen, CTD achieves statistically better, equivalent, and worse HV than the competing methods in 313, 118, and 9 out of 440 pairwise comparisons, respectively. These results further demonstrate the effectiveness of CTD across a broader range of optimisation problems.

W4: Too much text on page 9, and experiments are not strong.

Page 9 of the original paper discusses the results of the sensitivity analysis and the alternative acquisition function, followed by the conclusion. The detailed results are provided in the appendix (table 2, page 25).

Our experimental evaluation includes comparisons with well-established methods under noiseless, noisy, and high-dimensional settings, sensitivity analyses, ablation studies, experiments with alternative acquisition functions, and analyses of switching times and distances to the Pareto front. Regarding the experimental performance, across all 1,158 pairwise comparisons with well-established methods (Section 6.1), CTD achieves statistically better, equivalent, and worse HV in 811 (70.0%), 306 (26.4%), and 41 (3.5%) cases, respectively.

References

[1] Quality evaluation of solution sets in multiobjective optimisation: A survey, ACM Computing Surveys, 2019.

[2] Efficient multiobjective optimization employing Gaussian processes, spectral sampling and a genetic algorithm, JOGO, 2018.

[3] Differentiable expected hypervolume improvement for parallel multi-objective Bayesian optimization, in NeurIPS, 2020

[4] Pareto set learning for expensive multi-objective optimization, in NeurIPS, 2022

[5] Unexpected improvements to expected improvement for Bayesian optimization, in NeurIPS, 2023

[6] qPOTS: Efficient batch multiobjective Bayesian optimization via Pareto optimal Thompson sampling, in AISTATS, 2025

[7] Multi-objective Bayesian optimization via adaptive ϵ\epsilon-constraints decomposition, in ICML, 2026.

[8] Multi-objective and multi-fidelity Bayesian optimization of laser-plasma acceleration, Physical Review Research, 2023.

[9] Handbook of Writing for the Mathematical Sciences, SIAM, 1998.

[10] Regret for expected improvement over the best-observed value and stopping condition, In ACML, 2017.

[11] Using well-understood single-objective functions in multiobjective black-box optimization test suites, Evolutionary Computation, 2022.

[12] Real-parameter black-box optimization benchmarking 2009: Noiseless functions definitions, INRIA Report, 2009.

[13] BBOB instance analysis: Landscape properties and algorithm performance across problem instances, in EvoStar, 2022.

[14] Transfer of multi-objectively tuned CMA-ES parameters to a vehicle dynamics problem, in EMO, 2023.

[15] Pareto landscape: Visualising the landscape of multi-objective optimisation problems, in PPSN, 2024.

Reviewer_8rsGOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes converge-then-diversify (CTD), a two-stage framework for multi-objective Bayesian optimisation under limited evaluation budgets. CTD has two stages: convergence-focused search and diversity-focused search. The convergence-focused search adopts equal-weight EI based on the augmented Tchebycheff scalarisation. The default diversity-focused search adopts EHVI and the authors also consider ParEGO as an alternative choice. The paper also proposes how to decide when to switch from convergence-focused search to diversity-focused search. Experiments cover noiseless, noisy, and high-dimensional problems over established baselines.

Critical strengths
  1. The paper compares against several relevant methods, uses 30 independent runs and tests on multiple benchmarks.
  2. The experiments cover different settings: noiseless, noisy and high-dimensional cases. The sensitivity analysis on the threshold is also useful.
Critical weaknesses
  1. The writing is confusing and the main claim is unclear. The paper presents EHVI as the default acquisition function for the diversity stage, but concludes that random Tchebycheff scalarisations (ParEGO) can achieve even better performance. This leaves the recommended implementation unclear. If the scalarisation-based variant is the preferred implementation, it should receive greater prominence in the main experiments, with EHVI presented as an alternative. Otherwise, the authors should explain why EHVI remains the default despite the competitive performance of the scalarisation-based variant.

  2. I am also confused about CTD_{tch}. For this scalarisation-based CTD variant, both stages appear to use EI with augmented Tchebycheff scalarisation. Is the principal difference from ParEGO that CTD initially uses fixed equal weights and switches to random weights when the EI threshold is reached, whereas ParEGO uses random weights throughout? If so, this distinction should be stated explicitly. Comparisons against fixed-weight scalarised EI throughout would help establish the respective benefits of diversification.

  3. The results provide limited evidence for CTD’s general effectiveness. CTD achieves the best performance on only one of the six real-world benchmarks. On DTLZ1, DTLZ3, and DTLZ7 with m=3, it also underperforms at least one of ParEGO and EHVI, suggesting that the two-stage strategy does not consistently improve over its constituent methods.

  4. Most experiments involve only two or three objectives, with just one real-world benchmark involving four. Since the synthetic benchmarks readily support more objectives, could the authors evaluate CTD with m=4 and m=5 to assess its scalability?

  5. Missing baseline: The authors discuss STAGE-BO [1] in the related work but omit it from the experiments. Since STAGE-BO also promotes diversity across the Pareto front, it is a relevant baseline and should be included in the comparison.

[1] Yaohong Yang, Sammie Katt, and Samuel Kaski. Multi-objective Bayesian optimization via adaptive ε-constraints decomposition. ICML, 2026.

Questions for authors
  1. How often does the switch occur, and how close are solutions to the true front at that time? Could the authors provide these results to clarify how CTD works and why it performs better?

  2. In noisy experiments, is HV computed using latent noiseless objective values or noisy observations?

6 comments
AuthorsOpenReview ↗

Response [1/4]

W1&W2(a): Why is EHVI chosen as the default acquisition function for the diversity-focused stage in the main text rather than random Tchebycheff scalarisations (ParEGO), and how does scalarisation-based CTD differ from ParEGO?

The goal of this paper is to propose a new MOBO framework that decouples convergence and diversity into two separate stages, rather than considering them simultaneously throughout the search as existing MOBO methods such as EHVI and ParEGO do.

For the convergence-focused stage, we adopt EI based on the augmented Tchebycheff scalarisation, a widely used approach in MOBO [1-3]. In principle, any method (e.g., weighted sum) can be used provided that it can quickly guide the search towards a Pareto-optimal solution along a direction.

For the diversity-focused stage, we directly consider EHVI and random Tchebycheff scalarisations (ParEGO) as two instantiations, as they are widely used Pareto-based and scalarisation-based approaches. Note that in the diversity-focused stage, any method that accounts for both convergence and diversity can be instantiated, including most existing MOBO methods.

As such, given that EHVI and ParEGO are two instantiations of the second stage in the proposed framework, the presentation of their results is exchangeable (i.e., either in main text or in the appendices). We have made this issue clearer in Section 4.

W2(b): Comparison with fixed-weight scalarised EI throughout the search.

Taking the reviewer's suggestion, we have added a comparison with a variant that performs only the convergence stage until the evaluation budget is exhausted, denoted as CTD$_{converge}$. Table A below presents the results. As shown, CTD achieves statistically better HV on 16 of 20 problems and equivalent HV on the remaining 4. For these 4 cases (2- and 3-objective DTLZ1 and DTLZ3) that the two algorithms perform equally, the Pareto fronts are far from the initial solutions, making convergence more challenging. As a result, the switching criterion may not be met within the budget. We have now incorporated this comparison into the paper (Table 14, page 31).

Table A: HV obtained by CTD and its convergence-stage-only variant (CTD_converge).

MethodDTLZ1(2)DTLZ2(2)DTLZ3(2)DTLZ4(2)DTLZ5(2)DTLZ6(2)DTLZ7(2)
CTD_converge6.26e+00(1.2e+01)~1.82e-01(2.8e-02)+4.10e+02(7.5e+02)~1.83e-01(2.5e-02)+1.77e-01(2.1e-02)+2.75e-01(1.1e-01)+7.68e-02(7.9e-02)+
CTD6.68e+00(2.4e+01)3.46e-01(2.1e-02)4.17e+02(8.6e+02)3.52e-01(1.5e-02)3.49e-01(2.0e-02)3.19e-01(4.5e-02)1.71e-01(2.4e-02)
MethodDTLZ1(3)DTLZ2(3)DTLZ3(3)DTLZ4(3)DTLZ5(3)DTLZ6(3)DTLZ7(3)
CTD_converge4.25e+03(9.1e+03)~2.10e-01(3.6e-02)+1.54e+06(8.6e+05)~2.03e-01(3.0e-02)+4.02e-02(8.3e-03)+8.29e-02(3.3e-02)+1.65e-01(7.4e-05)+
CTD4.29e+03(5.3e+03)3.73e-01(8.0e-02)1.58e+06(8.9e+05)3.60e-01(9.0e-02)1.02e-01(1.3e-02)1.01e-01(1.7e-02)2.22e-01(1.9e-02)
MethodFour bar truss designPressure vessel designHatch cover designVehicle safetyCar side impactLPASummary(-/~/+)
CTD_converge4.05e+01(1.9e+00)+1.06e+10(2.3e+08)+2.09e+04(3.5e+02)+2.10e+01(2.0e+00)+1.97e+02(1.1e+01)+3.47e+07(3.5e+06)+0/4/16
CTD5.41e+01(2.8e-02)1.09e+10(1.8e+08)2.18e+04(2.0e+01)2.60e+01(6.4e-01)3.05e+02(1.5e+01)4.62e+07(3.7e+06)--/--/--

W3: CTD does not outperform competing methods on all problems. For example, it underperforms at least one of ParEGO and EHVI on DTLZ1 and DTLZ7 with m=3m=3.

Indeed, CTD can perform worse than ParEGO or EHVI on some problems, although on those problems it still outperforms several other baselines. In fact, across all 1,158 pairwise comparisons (comparison with well-established methods, Section 6.1), CTD performs statistically better, equivalently, and worse in 811 (70.0%), 306 (26.4%), and 41 (3.5%) cases, respectively. These results demonstrate that CTD can achieve very competitive performance, despite not being the best-performing one on every problem.

AuthorsOpenReview ↗

Response [2/4]

W4: Evaluation of CTD on DTLZ problems with m=4m=4 and m=5m=5.

CTD is expected to be competitive on many-objective problems. This is because CTD only considers convergence in the first stage, hence approaching the Pareto front substantially faster in a higher-dimensional space, whereas competing methods consider a number of directions simultaneously towards the Pareto front, resulting in slow convergence within a limited budget.

To confirm this, we have added the experiments on DTLZ problems with m=4m=4 and m=5m=5. Table B below presents the results. As shown, CTD achieves statistically better or equivalent HV than all competing methods on at least 13 out of the 14 problems. We have now incorporated this comparison into the paper (Table 3, page 26).

Table B: HV Results on DTLZ with m=4m=4 and m=5m=5.

MethodDTLZ1(4)DTLZ2(4)DTLZ3(4)DTLZ4(4)DTLZ5(4)DTLZ6(4)DTLZ7(4)
Sobol2.88e+05(3.9e+05)~1.47e+00(1.3e-01)+0.00e+00(0.0e+00)+5.43e-01(6.6e-02)+3.90e+00(9.2e-02)+1.91e+01(2.9e+00)+0.00e+00(0.0e+00)+
ParEGO1.06e+06(1.9e+06)~1.26e+00(3.3e-01)+9.06e+08(5.1e+08)~5.06e-01(2.9e-01)+4.56e+00(2.0e-01)+1.29e+02(7.4e+00)~3.23e-01(2.1e-01)~
TS-TCH1.74e+05(2.7e+05)~1.23e+00(2.4e-01)+0.00e+00(0.0e+00)+4.24e-01(8.4e-02)+4.19e+00(1.2e-01)+2.98e+01(3.8e+00)+0.00e+00(0.0e+00)+
PSL2.34e+06(8.5e+05)~1.35e+00(3.6e-01)+1.58e+09(3.7e+08)~5.30e-01(1.5e-01)+4.70e+00(9.8e-02)+1.33e+02(2.0e+00)~2.48e-01(1.0e-01)~
JES4.05e+05(9.1e+05)~1.13e+00(1.7e-01)+0.00e+00(0.0e+00)+3.40e-01(4.4e-02)+3.79e+00(9.1e-02)+2.56e+01(4.2e+00)+0.00e+00(0.0e+00)+
EHVI2.30e+06(4.6e+05)~1.18e+00(2.9e-01)+1.68e+09(2.2e+08)-3.19e-01(1.4e-01)+4.29e+00(1.4e-01)+1.33e+02(1.5e+00)~1.48e-01(3.1e-02)~
MORBO5.24e+05(5.9e+05)~1.35e+00(1.1e-01)+0.00e+00(0.0e+00)+3.85e-01(2.2e-01)+4.26e+00(1.9e-01)+3.10e+01(3.9e+00)+0.00e+00(0.0e+00)+
MOBO-OSD2.08e+06(6.9e+05)~9.63e-01(1.8e-01)+1.41e+09(4.6e+08)~5.85e-01(2.8e-01)+4.43e+00(3.9e-01)+1.34e+02(1.1e+00)~1.96e-01(3.4e-04)~
qPOTS3.09e+05(5.8e+05)~1.30e+00(2.2e-01)+0.00e+00(0.0e+00)+4.11e-01(1.2e-01)+4.07e+00(1.3e-01)+2.89e+01(3.3e+00)+0.00e+00(0.0e+00)+
CTD1.50e+06(1.9e+06)2.51e+00(1.3e-01)1.04e+09(6.3e+08)1.65e+00(1.7e-01)4.79e+00(1.6e-01)1.34e+02(2.7e+00)2.50e-01(1.5e-01)
MethodDTLZ1(5)DTLZ2(5)DTLZ3(5)DTLZ4(5)DTLZ5(5)DTLZ6(5)DTLZ7(5)
Sobol3.23e+07(4.7e+07)~3.98e-01(4.4e-02)+1.91e+10(4.3e+10)+5.30e-01(8.9e-02)+1.75e+01(4.1e-01)+3.82e+02(1.4e+01)+0.00e+00(0.0e+00)+
ParEGO1.33e+07(2.0e+07)~2.46e-01(1.1e-01)+2.18e+11(1.4e+11)+4.07e-01(2.1e-01)+2.02e+01(5.5e-01)~1.19e+03(8.9e+00)~3.80e-01(5.1e-02)+
TS-TCH3.00e+06(6.7e+06)+2.45e-01(1.3e-01)+0.00e+00(0.0e+00)+4.11e-01(1.3e-01)+1.86e+01(2.7e-01)+4.52e+02(2.8e+01)+0.00e+00(0.0e+00)+
PSL9.36e+07(8.9e+07)~3.68e-01(2.8e-01)+5.99e+11(7.9e+10)~5.20e-01(2.7e-01)+1.96e+01(1.3e+00)~1.19e+03(1.4e+01)~4.85e-01(1.1e-02)+
JES7.84e+06(1.5e+07)~2.46e-01(9.7e-02)+0.00e+00(0.0e+00)+3.04e-01(8.0e-02)+1.78e+01(4.8e-01)+3.97e+02(4.3e+01)+0.00e+00(0.0e+00)+
EHVI8.01e+07(5.4e+07)~3.26e-01(2.5e-01)+4.84e+11(9.4e+10)~3.47e-01(1.5e-01)+1.88e+01(7.8e-01)+1.20e+03(6.0e+00)~3.71e-01(4.2e-02)+
MORBO5.76e+06(9.1e+06)~3.99e-01(1.3e-01)+1.56e+10(3.5e+10)+4.74e-01(1.1e-01)+1.92e+01(3.6e-01)+4.35e+02(1.9e+01)+0.00e+00(0.0e+00)+
MOBO-OSD9.17e+07(6.4e+07)~4.65e-01(2.9e-01)+6.01e+11(7.2e+10)~4.20e-01(1.6e-01)+2.04e+01(4.0e-01)~1.18e+03(1.9e+01)~4.26e-01(6.8e-02)~
qPOTS3.09e+07(5.0e+07)~4.02e-01(2.2e-01)+0.00e+00(0.0e+00)+5.17e-01(1.3e-01)+1.89e+01(1.3e-01)+4.52e+02(2.7e+01)+0.00e+00(0.0e+00)+
CTD2.66e+07(2.2e+07)1.30e+00(1.5e-01)4.46e+11(1.9e+11)1.36e+00(1.8e-01)2.05e+01(5.5e-01)1.19e+03(1.7e+01)5.56e-01(2.4e-01)
AuthorsOpenReview ↗

Response [3/4]

W5: Missing comparison with STAGE-BO [Yang et al., 2026].

In this work, we propose an approach that decouples convergence from diversity, rather than considering both simultaneously throughout the search. To evaluate the proposed method, we considered nine representative MOBO methods. These methods cover representative classes of MOBO, i.e., scalarisation-based (EI-based ParEGO, TS-based TS-TCH, and UCB-based PSL) and Pareto-based (HV-based EHVI, Information-based JES, and maximin-distance-based qPOTS). Like these methods, STAGE-BO also considers convergence and diversity throughout the search, which uses a maximin-distance criterion to identify underexplored regions of the estimated Pareto front.

That said, taking the reviewer's suggestion, we have included this baseline in the experiments. Table C below presents the results. As shown, CTD achieves statistically better, equivalent, and worse HV than STAGE-BO on 11, 6, and 3 out of 20 problems, respectively. We have now incorporated this comparison into the paper (Table 1 and Figure 2 on page 7, and Figure 3 on page 8).

Table C: HV obtained by CTD and STAGE-BO.

MethodDTLZ1(2)DTLZ2(2)DTLZ3(2)DTLZ4(2)DTLZ5(2)DTLZ6(2)DTLZ7(2)
STAGE-BO0.00e+00(0.0e+00)~9.30e-03(1.1e-02)+0.00e+00(0.0e+00)~9.32e-03(1.1e-02)+9.29e-03(1.1e-02)+1.77e-01(5.8e-02)+6.74e-01(5.9e-02)-
CTD6.68e+00(2.4e+01)3.46e-01(2.1e-02)4.17e+02(8.6e+02)3.52e-01(1.5e-02)3.49e-01(2.0e-02)3.19e-01(4.5e-02)1.71e-01(2.4e-02)
MethodDTLZ1(3)DTLZ2(3)DTLZ3(3)DTLZ4(3)DTLZ5(3)DTLZ6(3)DTLZ7(3)
STAGE-BO3.33e+03(3.6e+03)+2.85e-02(1.5e-02)+1.02e+06(8.2e+05)+2.90e-02(1.5e-02)+1.72e-02(2.9e-02)+7.35e-02(1.8e-02)+8.17e-01(8.1e-02)-
CTD4.29e+03(5.3e+03)3.73e-01(8.0e-02)1.58e+06(8.9e+05)3.60e-01(9.0e-02)1.02e-01(1.3e-02)1.01e-01(1.7e-02)2.22e-01(1.9e-02)
MethodFour bar truss designPressure vessel designHatch cover designVehicle safetyCar side impactLPASummary(-/~/+)
STAGE-BO5.40e+01(2.8e-02)+1.09e+10(4.8e+07)~2.12e+04(7.6e+02)+2.65e+01(5.4e-01)~3.25e+02(1.8e+00)-3.95e+07(1.6e+06)~3/6/11
CTD5.41e+01(2.8e-02)1.09e+10(1.8e+08)2.18e+04(2.0e+01)2.60e+01(6.4e-01)3.05e+02(1.5e+01)4.62e+07(3.7e+06)--/--/--
Reviewer_GyKuOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

in this work, the authors propose a warm start method for multi-objective model-based blackbox optimization. a large body of literatures in this areas has focused on either coming up with a better acquisition function or designing a new optimization algorithm (often mixing in some stochasticity for exploration). the authors consider the problem of initialization.

the central thesis of this paper is that multi-objective optimization can be solved better if the process starts from a point on the pareto frontier and that it is easy to find one such point by focusing purely on a single (arbitrary) combination of objective functions (thereby ignoring the diversity in how they can be combined.) this is a reasonable assumption, and the authors use uniform weighting of different objective functions in their warm start phase (convergence phase, as they call it.) this is followed by the diversification phase, which is an actual multi-objective optimization phase.

the authors run an extensive set of evaluation and shows that in many cases, this strategy results in a greater hypervolume. there are some limited set of analyses, although i would appreciate more analyses, including (a) are the solutions found after the convergence phase actually differ significantly from the solution after the convergence phase? (that is, unless these solutions on the pareto frontier are well connected, is it possible that we are simply stuck in one solution?) (b) the effect of using different weighting scheme during the convergence phase, and (c) comparison against running many convergence phases in parallel only (this would be in fact similar to scalarization except that we take a long optimization trajectory for each scalarization strategy.)

in other words, it’s an interesting observation that initialization matters (or perhaps trivial for optimization researchers who have worked with any non-convex objective functions) and that this helps to a certain degree. it would be nice to understand precisely when and how it actually helps, beyond intuition.

Critical strengths

see my summary above.

Critical weaknesses

see my summary above.

Questions for authors

see my summary above.

3 comments
AuthorsOpenReview ↗

Response [1/3]

Q1&W1: Do the solutions found during the diversity-focused stage differ significantly from those found during the convergence-focused stage, e.g., on disconnected Pareto fronts?

In CTD, the convergence-focused stage aims to quickly approach a single Pareto-optimal solution, while the diversity-focused stage aims to spread solutions across the Pareto front. It is worth noting that both stages perform global acquisition optimisation rather than starting from evaluated solutions. Following BoTorch's default procedure, we generate 512 Sobol samples across the entire search space and apply L-BFGS from 10 promising starting points. Thus, the diversity-focused stage can explore entire search space, including disconnected ones, e.g., the pressure-vessel design problem.

To confirm this and address your concern, we added a convergence-stage-only variant, CTD$_{converge}$, which performs the convergence-focused stage until the budget is exhausted. Table A below presents the results. As shown, CTD achieves statistically better HV on 16 of 20 problems and equivalent HV on the remaining 4. For these 4 cases (2- and 3-objective DTLZ1 and DTLZ3), where the two algorithms perform equally, the Pareto fronts are far from the initial solutions, making convergence more challenging. As a result, the switching criterion may not be met within the budget.

We have now incorporated the above comparison into the paper (Table 14, page 31).

Table A: HV obtained by CTD and its convergence-stage-only variant (CTD_converge).

MethodDTLZ1(2)DTLZ2(2)DTLZ3(2)DTLZ4(2)DTLZ5(2)DTLZ6(2)DTLZ7(2)
CTD_converge6.26e+00(1.2e+01)~1.82e-01(2.8e-02)+4.10e+02(7.5e+02)~1.83e-01(2.5e-02)+1.77e-01(2.1e-02)+2.75e-01(1.1e-01)+7.68e-02(7.9e-02)+
CTD6.68e+00(2.4e+01)3.46e-01(2.1e-02)4.17e+02(8.6e+02)3.52e-01(1.5e-02)3.49e-01(2.0e-02)3.19e-01(4.5e-02)1.71e-01(2.4e-02)
MethodDTLZ1(3)DTLZ2(3)DTLZ3(3)DTLZ4(3)DTLZ5(3)DTLZ6(3)DTLZ7(3)
CTD_converge4.25e+03(9.1e+03)~2.10e-01(3.6e-02)+1.54e+06(8.6e+05)~2.03e-01(3.0e-02)+4.02e-02(8.3e-03)+8.29e-02(3.3e-02)+1.65e-01(7.4e-05)+
CTD4.29e+03(5.3e+03)3.73e-01(8.0e-02)1.58e+06(8.9e+05)3.60e-01(9.0e-02)1.02e-01(1.3e-02)1.01e-01(1.7e-02)2.22e-01(1.9e-02)
MethodFour bar truss designPressure vessel designHatch cover designVehicle safetyCar side impactLPASummary(-/~/+)
CTD_converge4.05e+01(1.9e+00)+1.06e+10(2.3e+08)+2.09e+04(3.5e+02)+2.10e+01(2.0e+00)+1.97e+02(1.1e+01)+3.47e+07(3.5e+06)+0/4/16
CTD5.41e+01(2.8e-02)1.09e+10(1.8e+08)2.18e+04(2.0e+01)2.60e+01(6.4e-01)3.05e+02(1.5e+01)4.62e+07(3.7e+06)--/--/--
AuthorsOpenReview ↗

Response [2/3]

Q2&W2: How do different weighting schemes affect the convergence-focused stage?

Good question. In CTD, we use equal weights during the convergence-focused stage to guide the search towards a balanced trade-off solution. As the reviewer points out, different weights may guide the search towards different regions of the Pareto front. To investigate this, we evaluated four alternative weighting schemes, including two with increasing weights and two with decreasing weights:

\frac{(1,\ldots,m)}{\sum_{i=1}^{m}i},\quad \frac{(m-1,\ldots,0)}{\sum_{i=0}^{m-1}i},\quad \frac{(m,\ldots,1)}{\sum_{i=1}^{m}i}.$$ For a bi-objective problem, these are $(0,1)$, $(1/3,2/3)$, $(1,0)$, and $(2/3,1/3)$. The switching criterion and diversity-focused stage remain unchanged. Table B below presents the results. As shown, the two non-extreme weighting schemes achieve performance comparable to the original CTD, with statistically equivalent HV on 19 and 20 out of the 20 problems, respectively. In contrast, CTD achieves better/equivalent/worse HV on 9/9/2 and 8/12/0 problems compared with the two extreme weighting schemes, respectively. These results suggest that CTD is robust to moderate changes in weights, while extreme weights may negatively affect performance. We have now incorporated the above results into the paper (Table 13, page 30). **Table B: HV obtained by CTD with five weighting schemes.** |**Method**|**DTLZ1(2)**|**DTLZ2(2)**|**DTLZ3(2)**|**DTLZ4(2)**|**DTLZ5(2)**|**DTLZ6(2)**|**DTLZ7(2)**| |-|-|-|-|-|-|-|-| |CTD_0-(m-1)|0.00e+00(0.0e+00)~|9.84e-02(8.7e-02)+|3.45e+02(6.2e+02)~|1.01e-01(8.1e-02)+|1.12e-01(8.3e-02)+|2.11e-01(2.9e-02)+|1.59e-01(1.4e-01)~| |CTD_1-m|2.25e-01(9.3e-01)~|3.42e-01(2.5e-02)~|6.38e+02(1.5e+03)~|**3.56e-01(1.4e-02)~**|3.40e-01(2.3e-02)~|2.98e-01(5.3e-02)~|1.70e-01(2.3e-02)~| |CTD_(m-1)-0|3.34e+00(1.7e+01)~|8.26e-02(5.8e-02)+|4.06e+02(5.2e+02)~|8.50e-02(7.4e-02)+|8.70e-02(6.9e-02)+|2.14e-01(3.6e-02)+|1.68e-01(2.3e-02)~| |CTD_m-1|6.22e-01(3.1e+00)~|3.36e-01(3.1e-02)~|**6.96e+02(9.9e+02)~**|3.48e-01(1.9e-02)~|3.40e-01(2.4e-02)~|3.13e-01(6.8e-02)~|**1.75e-01(2.8e-02)~**| |CTD|**6.68e+00(2.4e+01)**|**3.46e-01(2.1e-02)**|4.17e+02(8.6e+02)|3.52e-01(1.5e-02)|**3.49e-01(2.0e-02)**|**3.19e-01(4.5e-02)**|1.71e-01(2.4e-02)| |**Method**|**DTLZ1(3)**|**DTLZ2(3)**|**DTLZ3(3)**|**DTLZ4(3)**|**DTLZ5(3)**|**DTLZ6(3)**|**DTLZ7(3)**| |CTD_0-(m-1)|**9.28e+03(7.2e+03)~**|2.77e-02(4.9e-02)+|**1.64e+06(6.5e+05)~**|2.98e-02(3.4e-02)+|4.26e-03(9.7e-03)+|6.55e-02(1.4e-02)+|1.84e-01(1.2e-01)~| |CTD_1-m|5.78e+03(7.3e+03)~|3.42e-01(8.8e-02)~|1.63e+06(9.4e+05)~|2.99e-01(7.4e-02)~|9.41e-02(1.7e-02)~|9.89e-02(1.8e-02)~|**2.25e-01(1.2e-02)~**| |CTD_(m-1)-0|5.58e+03(6.1e+03)~|7.45e-02(7.9e-02)+|1.63e+06(6.4e+05)~|6.57e-02(7.1e-02)+|4.32e-02(3.2e-02)+|6.41e-02(1.3e-02)+|2.23e-01(1.7e-02)~| |CTD_m-1|1.47e+03(2.6e+03)~|3.59e-01(6.8e-02)~|1.37e+06(8.8e+05)~|3.57e-01(6.3e-02)~|9.60e-02(2.0e-02)~|8.74e-02(2.3e-02)~|2.23e-01(1.7e-02)~| |CTD|4.29e+03(5.3e+03)|**3.73e-01(8.0e-02)**|1.58e+06(8.9e+05)|**3.60e-01(9.0e-02)**|**1.02e-01(1.3e-02)**|**1.01e-01(1.7e-02)**|2.22e-01(1.9e-02)| |**Method**|**Four bar truss design**|**Pressure vessel design**|**Hatch cover design**|**Vehicle safety**|**Car side impact**|**LPA**|**Summary(-/~/+)**| |CTD_0-(m-1)|5.41e+01(4.4e-02)~|1.06e+10(4.7e+08)+|2.08e+04(2.5e+03)~|**2.64e+01(1.1e+00)-**|**3.23e+02(8.1e+00)-**|4.31e+07(5.6e+06)~|2/9/9| |CTD_1-m|**5.41e+01(3.1e-02)~**|1.09e+10(1.3e+08)~|2.18e+04(2.2e+01)~|2.55e+01(1.6e+00)~|2.76e+02(1.0e+02)~|4.47e+07(4.5e+06)~|0/20/0| |CTD_(m-1)-0|5.41e+01(4.6e-02)~|**1.09e+10(6.4e+07)~**|**2.18e+04(3.0e+01)~**|2.57e+01(4.0e-01)~|3.12e+02(1.1e+01)~|4.44e+07(4.8e+06)~|0/12/8| |CTD_m-1|5.40e+01(4.4e-02)+|1.08e+10(1.9e+08)~|2.18e+04(2.0e+01)~|2.60e+01(8.5e-01)~|3.13e+02(1.3e+01)~|**4.63e+07(4.2e+06)~**|0/19/1| |CTD|5.41e+01(2.8e-02)|1.09e+10(1.8e+08)|2.18e+04(2.0e+01)|2.60e+01(6.4e-01)|3.05e+02(1.5e+01)|4.62e+07(3.7e+06)|--/--/--|
AuthorsOpenReview ↗

Response [3/3]

Q3&W3: Comparison against running multiple convergence phases with different scalarisation weights.

This is an interesting idea. We expect that running multiple convergence-focused searches with different weights could guide the search towards different regions of the Pareto front. However, under a limited evaluation budget, using many weights can prevent the search from sufficiently converging in each direction. Consequently, the obtained solutions may exhibit good diversity but remain far from the Pareto front.

To confirm this, we added this variant, where multiple convergence-focused searches were performed independently using different fixed weights. Specifically, for an mm-objective problem, we used m+1m+1 weight vectors: mm focusing on individual objectives and one using equal weights. For example, for bi-objective problems, we used (1,0)(1,0), (0,1)(0,1), and (0.5,0.5)(0.5,0.5). We applied the same strategy to three- and four-objective problems.

Table C below presents the results. As shown, CTD achieves statistically better, equivalent, and worse HV than this variant on 15, 5, and 0 out of 20 problems, respectively. These results suggest that first-converging-then-diversifying is more effective under a limited evaluation budget. We have now incorporated this comparison into the revised paper (Table 15, page 31).

Table C: HV obtained by CTD and Multi_CTD$_{converge}$.

MethodDTLZ1(2)DTLZ2(2)DTLZ3(2)DTLZ4(2)DTLZ5(2)DTLZ6(2)DTLZ7(2)
Multi_CTD_converge0.00e+00(0.0e+00)~1.96e-02(2.3e-02)+9.27e+01(4.0e+02)~3.14e-02(4.3e-02)+2.09e-02(3.7e-02)+9.13e-02(5.0e-02)+1.98e-01(1.5e-01)~
CTD6.68e+00(2.4e+01)3.46e-01(2.1e-02)4.17e+02(8.6e+02)3.52e-01(1.5e-02)3.49e-01(2.0e-02)3.19e-01(4.5e-02)1.71e-01(2.4e-02)
MethodDTLZ1(3)DTLZ2(3)DTLZ3(3)DTLZ4(3)DTLZ5(3)DTLZ6(3)DTLZ7(3)
Multi_CTD_converge3.49e+03(4.4e+03)~2.00e-02(2.5e-02)+1.12e+06(8.8e+05)~2.84e-02(3.4e-02)+1.33e-03(4.4e-03)+5.08e-02(2.3e-02)+5.12e-02(1.1e-01)+
CTD4.29e+03(5.3e+03)3.73e-01(8.0e-02)1.58e+06(8.9e+05)3.60e-01(9.0e-02)1.02e-01(1.3e-02)1.01e-01(1.7e-02)2.22e-01(1.9e-02)
MethodFour bar truss designPressure vessel designHatch cover designVehicle safetyCar side impactLPASummary(-/~/+)
Multi_CTD_converge4.50e+01(1.8e+00)+1.08e+10(1.2e+08)+2.12e+04(1.7e+02)+2.54e+01(4.0e-01)+2.55e+02(8.4e+00)+3.40e+07(4.1e+06)+0/5/15
CTD5.41e+01(2.8e-02)1.09e+10(1.8e+08)2.18e+04(2.0e+01)2.60e+01(6.4e-01)3.05e+02(1.5e+01)4.62e+07(3.7e+06)--/--/--

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.