acceptodds
Under review as a conference paper at ICLR 2027

Learning What to Evaluate: Correlation-Aware Decoupling for Multiobjective Bayesian Optimization

Abstract

Multiobjective Bayesian optimization (MOBO) with Gaussian process (GP) surrogates is a sample efficient approach to solving multiobjective optimization problems. In MOBO, a Bayesian decision theoretic acquisition function guides the adaptive selection of new candidate inputs, on which objectives and constraints are evaluated to update the surrogate model sequentially. Existing approaches maintain independent GP models for the objectives and constraints, with new observations evaluating all objectives and constraints in a coupled fashion. However, the objectives and constraints often contain inherent correlations which, if exploited, can enable decoupled evaluations where only a subset of them are evaluated at each round. We present a new approach that leverages a multitask GP model to jointly learn all objectives and constraints, and propose a total correlation metric that enables identifying an optimal subset of objectives and constraints to be evaluated at every round, even under uniform evaluation costs. Theoretically, we show that our acquisition policy is asymptotically consistent despite decoupling and that our proposed decoupled subset selection rule maximizes the expected posterior entropy reduction about unevaluated tasks under mild conditions. Empirically, we show that our approach outperforms coupled and decoupled baselines in the state of the art.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Bayesian methods

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_nkzQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors decouple evaluations for MOBO and identify a subset of objectives and constraints to be evaluated every round. Their approach outperforms existing baselines across synthetic and real-world problems.

Critical strengths
  • The problem is well-motivated, and the solution seems sensible and effective. Treating the constraints as part of the optimization procedure seems important.
  • The empirical results also seem promising, and the method outperforms many existing baselines across a diverse set of problems, including synthetic tasks, real tasks, with and without constraints, etc.
Critical weaknesses
  • The benefits of the proposed mechanism are not demonstrated clearly. The differences between qPOTS-DOE and the existing benchmarks, especially HVKG which the authors describe as the main competitor, are not explained in the related work and instead require detailed reading of the corresponding papers. In addition, many of the existing competitive methods are single-task GPs, and it would be useful to demonstrate the benefits of the MTGP approach explicitly by comparing it to existing MTGP methods.
  • The benefits of the two-stage decoupling mechanism are also unclear, and it would be interesting to see the contributions of different components of this setup, such as the impact of including the constraints. It could also be useful to evaluate the benefits of the correlation metric more directly, perhaps on different synthetic datasets. For the benefits of the optimal subset, it may be interesting to compare the optimal subset algorithm with alternatives.

I'm open to updating my score if the reasons for the improved results are demonstrated more clearly.

Questions for authors
  • Have you tried ablating the two stages of the algorithm? How robust is it to different hypers, like the number of objectives selected?
  • Do qLogNEHVI baselines also use the MTGP?
1 comment

Dear Reviewer nkzQ,

As the discussion phase comes to an end, please acknowledge the author's response and consider leaving them a comment on whether it has resolved some or all of your concerns.

Thanks!

Reviewer_oo38OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes qPOTS-DOE, a multi-objective Bayesian optimization algorithm for settings where objectives and constraints can be evaluated independently. Building on qPOTS, the method fits a multitask GP, draws a joint Thompson sample of all objectives and constraints, and uses NSGA-II to approximate the sampled feasible Pareto set. It then selects q candidates using a distance-based rule to promote diversity. The main contribution is a two-stage decoupling mechanism: at each candidate, a total-correlation score determines whether to evaluate all tasks or only a subset; if decoupling is selected, a mutual-information criterion chooses M tasks that are most informative about the remaining tasks. The theoretical analysis establishes posterior consistency under specified assumptions and the optimality of subset selection for a fixed size in terms of local expected entropy reduction. Experiments on eight synthetic and four real-world benchmark problems demonstrate promising evaluation efficiency relative to competing methods.

Critical strengths
  • Relevant practical problem: The paper addresses decoupled multi-objective optimization, which has received less attention than the coupled setting. When objectives and constraints can be evaluated separately, selectively evaluating them can avoid unnecessary evaluations.
  • Intuitive and well-motivated framework: The total-correlation threshold and mutual-information-based subset selection provide a clear mechanism for exploiting task dependencies to reduce redundant evaluations.
  • Promising empirical results: Experiments across synthetic and real-world benchmarks suggest improved evaluation efficiency relative to coupled and decoupled baselines.
Critical weaknesses
  1. Unclear necessity of the decoupling threshold: Figure 5 shows marginal changes in performance over a wide range of τ\tau, raising questions about the practical benefit of the TC gate.
  2. Computational cost and scalability: Figure 15 reports approximately 4×1044\times10^4 seconds ~ 11 hours for 100 iterations on OSY. This is fine given small problems, but MOBO problems tend to scale quickly with number of dimensions, number of iterations, number of objectives, hence the method can incur a large overhead in practice.
  3. Limited methodological novelty: The framework builds directly on qPOTS, retaining posterior Thompson sampling, NSGA-II-based Pareto-set approximation, and distance-based batch selection. MTGPs, total correlation, and Gaussian mutual information are established techniques. The methodological contribution therefore primarily lies in integrating these components into a decoupled evaluation policy.
  4. Unclear hypervolume computation under partial observations: Computing hypervolume requires complete objective vectors. However, the paper does not clearly explain how missing objective and constraint values are obtained for partially evaluated candidates. This makes comparison against coupled methods less clear.
Questions for authors
  1. Given the limited sensitivity to τ\tau in Figure 5, how much benefit does the TC gate provide compared with an always-decoupling strategy (performance and runtime)? Also, could a coupled evaluation (M=K+C) instead be incorporated as an option when selecting M? This would add only one subset to evaluate.
  2. In Section 3.2, page 6, paragraph “Approximate maximization”, for the sentence “Then, we repeat Equation (8) … smallest f_X(S)”, should “smallest” be “largest”, since we want to maximize information gain?\
  3. How is the proposed information-gain-based subset selection compared to other strategies, e.g., random selection?
  4. Could the authors also report hypervolume against number of iterations? This would help distinguish improvements against coupled methods.
Reviewer_NEaZOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes qPOTS-DOE, a method for reducing the number of evaluations in multi-objective Bayesian optimization (MOBO) by leveraging the correlation between objective and constraint functions. It models each function simultaneously using Multitask GP (MTGP), selects evaluation points via Thompson Sampling, and then determines a subset of functions to evaluate based on Total Correlation (TC) and Mutual Information (MI). Theoretically, it demonstrates the posterior consistency of MTGP under certain conditions and the optimality of the information gain for unevaluated functions. On benchmarks derived from synthetic and real-world problems, it achieves higher hypervolume with fewer evaluations than existing methods in many cases.

Critical strengths

The idea of correlation-based partial evaluation is clear and practically relevant. While conventional Decoupled MOBO primarily leveraged differences in evaluation costs between functions, this study utilizes correlations between functions to create the benefits of partial evaluation even when evaluation costs are equal. Furthermore, the ability to handle not only objective functions but also constraint functions is significant.

The extensive empirical evaluation is another strength. The method consistently demonstrates strong results across multiple benchmarks derived from both synthetic and real-world problems. Comparisons with qPOTS using MTGP, as well as various ablation studies, provide substantial evidence for the improvement in sample efficiency.

Critical weaknesses

Function selection based on MI does not directly correspond to the objective of MOBO. While the proposed method is expected to contribute to the estimation of objective and constraint functions by maximizing the amount of information gained about unevaluated functions, this does not necessarily mean it selects functions that contribute to improving the estimated Pareto frontier or hypervolume. There is a lack of analysis to support the validity of this selection criterion, as well as comparisons with other subset selection strategies.

Theoretical guarantees do not sufficiently support the performance of the proposed algorithm. Theorem 4.1 establishes posterior consistency of MTGP under the assumption that each function is sufficiently densely observed across the entire input space; it is not guaranteed that this condition is satisfied by the actual evaluation policy proposed. Furthermore, the optimality in Theorem 4.2 pertains to local information acquisition and does not directly guarantee the optimization performance of MOBO. Consequently, the theoretical contribution is limited.

Improvements in overall optimization efficiency, accounting for computational cost, have not been sufficiently validated. While the experiments evaluate hypervolume relative to the number of function evaluations, the proposed method performs fewer function evaluations per iteration; therefore, for the same number of function evaluations, it may execute more GP updates and acquisition point searches. An evaluation of total computational time is therefore necessary. Although a comparison of execution times is provided in the Supplementary material, there is no comparison of computational time within the same function evaluation budget or of hypervolume relative to total execution time, making the practical efficiency improvements unclear.

Questions for authors

Q1. Can the authors demonstrate that MI-based function selection improves optimization performance by comparing it with random subset selection using the same MTGP? Additionally, how is the number of functions to be evaluated (MM) determined in the experiments? While this method maximizes MI, the paper states that MM is selected by minimizing MI, which seems inconsistent with the stated objective.

Q2. Can the hypervolume be compared relative to the cumulative computation time for the same number of function evaluations, or relative to the total execution time including function evaluation costs? In particular, please clarify how expensive function evaluations need to be for the proposed method to become advantageous.

1 comment

Dear Reviewer NEaZ,

As the discussion phase comes to an end, please acknowledge the author's response and consider leaving them a comment on whether it has resolved some or all of your concerns.

Thanks!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.