Converge Then Diversify: Decoupling Convergence and Diversity in Multi-Objective Bayesian Optimisation
Chao Jiang, Yueling Huang, Miqing Li
cs.AI · stat.ML
Quick verdict
Proposes converge-then-diversify MOBO with an EI-based switch; simple idea backed by 446 comparisons, but per-run switching behaviour remains under-reported.
Summary
The paper proposes converge-then-diversify (CTD), a two-stage framework for multi-objective Bayesian optimisation (MOBO) under very tight evaluation budgets. Instead of pursuing convergence and diversity simultaneously, CTD first drives the search to a single Pareto-front point using expected improvement (EI) over a fixed equal-weight augmented Tchebycheff scalarisation, then switches to a diversity-focused stage with EHVI (or ParEGO-style random scalarisations as an alternative) once max EI falls below κ=10⁻⁴, reusing a published EI stopping rule. Experiments cover 20 benchmark (DTLZ1–7, m=2/3) and real-world problems under noiseless, 10%-noise, and d=50 high-dimensional settings against nine baselines (EHVI/NEHVI, ParEGO/NParEGO, TS-TCH, PSL, JES, MORBO, MOBO-OSD, qPOTS, Sobol), 30 runs each, with Wilcoxon rank-sum tests and Holm-Bonferroni correction. Across 446 pairwise comparisons CTD wins 72.9%, ties 21.1%, loses 6.1%, with the largest margins under tight budgets and in high dimensions. Limitations: weak performance on disconnected fronts (statistically worse than four methods on noiseless DTLZ7, m=2), no per-run statistics on whether/when the stage switch fires, κ's scale-sensitivity only partially addressed, and no code release stated. The contribution is a scheduling insight built from standard components rather than a new acquisition function.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 2
● low 4
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Decoupling convergence and diversity into two sequential stages (converge via fixed equal-weight augmented Tchebycheff EI, then diversify via EHVI) improves MOBO performance under very tight evaluation budgets.
- An EI-based stopping rule (max EI < 10⁻⁴, log-scale for LogEI) is an effective stage-switch criterion, stable for κ between 10⁻³ and 10⁻⁵.
- Across 446 pairwise comparisons (180 noiseless, 140 noisy, 126 high-dimensional at d=50) on 20 problems, CTD statistically outperforms nine baselines in 72.9% of cases, ties in 21.1%, and loses in only 6.1%.
- The framework generalises across stage-2 instantiations: ParEGO-style random scalarisations (CTDtch) remain competitive and improve DTLZ7 (m=2, noiseless) from HV 1.71e-01 to 3.75e-01.
- Convergence-first scheduling is especially effective in high dimensions (d=50), where CTD beats the dedicated high-dimensional method MORBO on all 14 problems.
Strengths
Headline statistics reconcile exactly with per-table evidence
The abstract's 446 comparisons (72.9%/21.1%/6.1%) match the per-setting sign counts: 129/34/17 of 180 noiseless (Table 1), 85/52/3 of 140 noisy (Table 2), 111/8/7 of 126 at d=50 (Table 7) — unusually auditable for a MOBO paper.
Broad, well-instrumented benchmarking
20 problems (DTLZ1–7 at m=2/3 plus six real-world designs incl. the LPA surrogate), three regimes (noiseless, 10% noise, d=50), nine baselines spanning HV-, information-, scalarisation- and TS-based families; log-acquisition variants applied uniformly per Ament et al. [29]; 30 runs per cell.
Genuine sensitivity and ablation work
Honest negative reporting
Timely high-dimensional angle
Weaknesses
Stage-switching behaviour never reported per run
Clear losses on disconnected fronts
High-variance HV presentation on DTLZ1/DTLZ3
Presentation defects that impede audit
No code or data availability statement
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets multi-objective Bayesian optimisation (MOBO) under very tight evaluation budgets and argues that pursuing convergence and diversity simultaneously is wasteful there. It proposes CTD: a convergence-focused stage (LogEI on a fixed equal-weight augmented Tchebycheff scalarisation) followed by a diversity-focused stage (LogEHVI, or ParEGO-style random scalarisations in the CTDtch variant), switching when max EI < κ = 10⁻⁴ per a published EI stopping rule. Central claims: (1) across 446 pairwise comparisons CTD wins 72.9%, ties 21.1%, loses 6.1%; (2) the advantage grows under tight budgets and at d=50; (3) the framework generalises across instantiations. The counts reconcile exactly with the per-setting summaries (129/34/17 of 180 noiseless, Table 1; 85/52/3 of 140 noisy, Table 2; 111/8/7 of 126 high-dimensional, Table 7) — unusually auditable. Note that 54 of the 446 comparisons are against the Sobol random baseline; excluding them the win rate is 278/392 ≈ 70.9%.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the converge-then-diversify (CTD) approach in multi-objective Bayesian optimisation?
CTD is a two-stage MOBO strategy for very tight evaluation budgets. It first optimises a single fixed scalarisation (equal-weight augmented Tchebycheff) with expected improvement to converge quickly to one point of the Pareto front, then switches to a hypervolume-improvement acquisition (EHVI) that spreads solutions along the front. The rationale is that chasing the whole front simultaneously wastes samples when very few evaluations are available.
How does CTD decide when to stop converging and start diversifying?
It monitors the maximum value of the scalarised expected improvement; once it drops below a threshold κ = 10⁻⁴ (applied on the log scale because LogEI is implemented), following a published EI-based stopping rule. A sensitivity study over κ ∈ {10⁻¹…10⁻⁶} shows results are statistically equivalent for κ between 10⁻³ and 10⁻⁵ on 19 of 20 problems.
How well does CTD perform against other MOBO methods?
On 20 benchmark (DTLZ1–7, 2–3 objectives) and real-world problems, under noiseless, 10%-noise and 50-dimensional settings, across 446 pairwise comparisons with 30 runs each, CTD was statistically better than the nine compared methods (EHVI/NEHVI, ParEGO/NParEGO, TS-TCH, PSL, JES, MORBO, MOBO-OSD, qPOTS, plus Sobol) in 72.9% of comparisons, equal in 21.1%, worse in 6.1% (Wilcoxon rank-sum with Holm-Bonferroni). Its weakest area is disconnected Pareto fronts such as DTLZ7.
What are the main limitations of CTD?
It is weakest on problems with disconnected Pareto fronts: on noiseless DTLZ7 (2 objectives) it is statistically worse than MOBO-OSD, ParEGO, PSL and JES, and swapping the diversity-stage acquisition recovers only part of the gap. The paper also does not report per-run statistics on whether or when the stage switch fires, κ is an absolute threshold whose scale-sensitivity is only partially explored, and no public code release is stated.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–4 are known only through captions and surrounding text; visual claims (solution placements in Figure 3, trajectory shapes in Figure 2) could not be independently verified from the extracted text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.