acceptodds
Under review as a conference paper at ICLR 2027

On Setting the Reference Point for the Hypervolume Indicator in Multi-Objective Bayesian Optimisation

Abstract

The Hypervolume (HV) indicator is widely used in Bayesian optimisation for expensive multi-objective optimisation problems, such as in Expected Hypervolume Improvement (EHVI). A key issue in HV-based methods is the specification of the reference point, which critically influences optimisation performance. In situations where an estimate of the problem's true Pareto front is unavailable (typically the case in real-world applications), a common strategy is to dynamically update the reference point according to the boundaries of the observed Pareto front. However, different regions of the Pareto front often exhibit different levels of reachability during the optimisation process. For example, certain boundary regions may be particularly difficult to discover. As such, selecting the reference point dynamically based on the currently observed Pareto front may bias the search and hinder the identification of the entire Pareto front, as solutions beyond the reference point contribute (approximately) zero to the EHVI/HV value. In this paper, we propose a simple method to address this issue. Specifically, we adaptively adjust the HV reference point to shift the search focus between the boundary and interior regions of the Pareto front, aiming to identify the boundaries first and then the interior regions. We validate the proposed method on a diverse set of 69 benchmark and real-world problems. Extensive experimental results demonstrate that it consistently outperforms various reference-point specification strategies.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Bayesian optimisation, multi-objective optimisation, expected hypervolume improvement, reference point specification

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_k8G6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies proper reference point specification in multi-objective bayesian optimization using expected hypervolume improvement. It propose two stage approach for this problem. In the first stage, the reference point is set iteratively per objective such that the optimization is focused on each of the objective iteratively (such as single objective optimization). In the 2nd stage, the reference point in the standard way, assuming the first stage has identified all the boundary point in the Pareto front. The paper has compared exhaustively with many different reference point specification approach and demonstrate the algorithm works empirically competitive.

Critical strengths
  • The paper propose a simple but empirically competitive method for reference point specification.
  • The paper is well-written and easy to follow.
Critical weaknesses
  • One of the main limitation of this paper is its rather practical, although has conduct extensive empirical investigation, the method itself is less sound; Additionally, the two step transfer also introduce the hyperparametr that can make the algorithm cumbersome to tune.
  • Another issue is that, different objective function may have different optimization difficulty, it is not sound by simply looping each objective with the hope that each objective can be optimized evenly well in the first stage.
Questions for authors
  • Following the 2nd weakness, could the author clarify if different objective function have different optimization difficulty, does switch of reference point to preference each objective function is appropriate or maybe more appropriate approach could be deviced?
  • Does the author feel the following paper [1] relevant to discuss in the paper?

[1] Zheng, R., & Wang, Z. (2024). Boundary decomposition for nadir objective vector estimation. Advances in Neural Information Processing Systems, 37, 14349-14385.

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

The authors focus on Expected Hypervolume Improvement (EHVI) in multi-objective Bayesian optimisation, and consider how to achieve a set of well-distributed Pareto solutions by adapting the reference points for EHVI. Specifically, the authors propose a two-phase reference point adaptation strategy. The first phase adjusts the reference point to encourage the discovery of boundary solutions for individual objectives, while the second phase switches to a more balanced reference point to explore the intermediate regions of the Pareto front. Experiments on 69 benchmark and engineering optimisation problems demonstrate improved hypervolume performance compared with two commonly used dynamic reference point strategies.

Critical strengths
  1. The paper addresses an important yet often overlooked aspect of hypervolume-based multi-objective Bayesian optimisation. The authors clearly demonstrate how different reference point specifications can influence the EHVI landscape and potentially restrict the exploration of the Pareto front.
  2. The proposed two-phase strategy is straightforward to implement and can be integrated into existing EHVI-based optimisation methods without modifying their surrogate models or acquisition functions.
Critical weaknesses
  1. The novelty is my major concern. The general idea of adjusting reference points or normalization bounds to improve the exploration of different Pareto front regions has already been studied extensively.
  2. The proposed method switches between objectives after a predefined number of consecutive iterations without hypervolume improvement and enters the second phase after a longer stagnation period. However, the absence of recent hypervolume improvement does not necessarily indicate that the corresponding Pareto front boundary has been adequately explored.
  3. The main experimental evaluation compares the proposed method primarily with two relatively simple dynamic reference point specifications.
Questions for authors
  1. Section 3.1 suggests that conventional EHVI fails partly because L-BFGS-B converges to suboptimal regions of the acquisition landscape. Could the authors clarify how they distinguish this optimisation issue from the limitations of the reference point itself?
  2. Could the authors clarify how the switching mechanism behaves when some objectives are substantially more difficult to optimise than others?
  3. Based on the existing experimental results, could the authors clarify the individual roles of boundary-focused exploration and subsequent balanced search?
Reviewer_m9G4OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes a two-phase strategy for setting the reference point in Expected Hypervolume Improvement (EHVI) for multi-objective Bayesian optimization (MOBO). In Phase 1, the reference point is chosen to emphasize one objective at a time using Eq. (1), cycling through the objectives according to a stagnation-based switching rule. In Phase 2, the method adopts a conventional reference point – the “1.1 times the observed Pareto front”. The main methodological contributions are the reference point formula in Eq. (1) and the switching mechanism. The experiments compare the proposed strategy with conventional reference point strategies and a fixed reference point based on the true Pareto front, all within a EHVI implementation. A sensitivity analysis examines the effect of the stagnation threshold used for switching.

Critical strengths
  • The proposed method is straightforward and easy to implement in practice.

  • The central idea is interesting: allocating part of the evaluation budget to focusing on individual objectives in Phase 1 before seeking a balanced approximation of the Pareto front in Phase 2.

  • The main experiments demonstrate improved performance over conventional reference point strategies across a broad set of benchmark problems.

Critical weaknesses
  1. Limited methodological contribution. The contribution mainly consists of a simple reference point formula, Eq. (1), and a stagnation-based switching mechanism, both described in Section 3.2. This makes me question the significance and contribution of the proposed method.

  2. Narrow experimental scope. Despite covering many benchmark problems, the evaluation appears restricted to noiseless, sequential optimization (these settings are not explicitly stated in the manuscript). Experiments in batch, noisy, or constrained settings would strengthen the evidence within the paper’s current limited scope.

  3. Unclear generality beyond EHVI. The proposed strategy is not connected to EHVI, yet it is evaluated only within EHVI implementation. Other HV-based acquisition functions, such as HVI evaluated from the posterior mean, could potentially benefit from this strategy. Potentially, the proposed method could be applied to other fields such as Evolutionary Algorithms [1].

  4. Insufficient justification for Eq. (1). The paper does not explain why the offset magnitude should be divided by the initial sample size (N).

  5. Limitations of the switching mechanism. The rule does not guarantee that Phase 2 is reached within the evaluation budget, since any HV improvement resets the stagnation counters. Furthermore, switching to Phase 2 requires msms consecutive evaluations without improvement, which grows linearly with the number of objectives and may consume a substantial portion of the budget. The experimental results mostly consist of bi-objective evaluation, which may hide this scalability issue.

  6. Lack of comparisons with other MOBO methods. The experiments compare reference-point strategies within EHVI but do not establish the resulting method’s competitiveness against other MOBO approaches. Though it is not directly related to the main contribution, comparisons with representative methods such as qParEGO [2], JES [3], or DGEMO [4] would help strength the practical significance of the improvements.

  7. Limited reporting of convergence behavior. The experimental results do not show the conventional performance plots, e.g., HV trajectories across iterations (similar to the second and forth plots in Figure 6). Compared to Table 1, which only display the final values, HV trajectories would better demonstrate performance throughout the evaluation budget and clarify the trade-off between spending evaluations in Phase 1 and achieving subsequent gains in Phase 2.

  8. Incomplete experimental reporting. Many details needed for reproducibility are missing or unclear, including problem dimensions and objective counts, the noise setting, GP likelihood and hyperparameter settings or constraints, and batch size.

[1] Nicola Beume, Boris Naujoks, and Michael Emmerich. Sms-emoa: multiobjective selection based on dominated hypervolume. European Journal of Operational Research, 181(3):1653–1669, 2007.

[2] Joshua Knowles. Parego: A hybrid algorithm with on-line landscape approximation for expensive multiobjective optimization problems. IEEE transactions on evolutionary computation, 2006.

[3] Ben Tu, Axel Gandy, Nikolas Kantas, and Behrang Shafei. Joint entropy search for multiobjective bayesian optimization. Advances in Neural Information Processing Systems, 35, 2022.

[4] Mina Konakovic Lukovic, Yunsheng Tian, and Wojciech Matusik. Diversity-guided multiobjective bayesian optimization with batch evaluations. Advances in Neural Information Processing Systems, 33, 2020.

Questions for authors
  1. Why is the proposed strategy specifically framed for EHVI? Could it also benefit other HV-based acquisition functions, such as HVI evaluated at the posterior mean, or other fields such as HV-based multi-objective evolutionary algorithms?

  2. How would the strategy extend to noisy, constrained, or batch optimization settings? As I mentioned, extending to one of these settings would greatly improve the contribution.

  3. What motivates Eq. (1), e.g., dividing the offset terms by the initial sample size (N)? How sensitive is the method to the number initial data points?

  4. Could Phase 1 instead use single-objective BO to optimize each objective in turn before switching to Phase 2? Single-objective optimization would be less computational expensive that EHVI-based optimization for phase 1.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.