MF-SCBO : Multi-fidelity Scalable Constrained Bayesian Optimization
Lucas Palazzolo, Mickaël Binois, Laëtitia Giraldi
cs.LG · math.OC
Quick verdict
MF-SCBO extends SCBO to multi-fidelity with an LOO-guarded trust region; strong gains on 6–100D constrained benchmarks, though SCBO wins some cases.
Summary
The paper proposes MF-SCBO, a multi-fidelity extension of Scalable Constrained Bayesian Optimization (SCBO) for expensive black-box problems that are simultaneously high-dimensional, constrained, and offer several fidelity levels with non-nested designs. Building on autoregressive co-kriging (with predictive-mean imputation for non-nested data), the method runs Thompson sampling inside an adaptive trust region and assigns a fidelity to each candidate via an uncertainty-per-cost criterion (Eq. 16). Two trust-region-center variants are introduced: 'evaluated', using only high-fidelity evaluations, and 'predicted', using the highest-fidelity posterior mean over all data, safeguarded by an analytical leave-one-out test (Eqs. 8–15) that forces a high-fidelity evaluation when predictions look unreliable. Constraints are assumed fidelity-independent. On synthetic benchmarks (Hartmann6D, Borehole8D, Ackley20D, Rosenbrock40D/100D) and realistic problems (solar thermal plant, XFOIL airfoil; 6–100 dimensions), MF-SCBO generally converges faster than SCBO and MF-MES, the latter becoming impractical beyond roughly 8 dimensions. The predicted variant wins when predictions are reliable (Ackley20D: 0.50 vs 9.40 for SCBO), while the evaluated variant is safer on Rosenbrock and solar7. Limitations: no theoretical guarantees; SCBO attains better final values on Rosenbrock; the sole multi-fidelity baseline is a penalized, time-limited MF-MES; results are reported as means ± 1.96 SD without significance tests; code is released in BoTorch.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
4 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
- MF-SCBO is, to the authors' knowledge, the first method to simultaneously handle high dimensionality, black-box constraints, an arbitrary number of fidelity levels, and non-nested sampling.
- The analytical leave-one-out criterion (Eqs. 8–15) keeps prediction-based trust-region centers reliable at negligible computational cost, without refitting any GP.
- MF-SCBO generally achieves better convergence than single-fidelity SCBO and MF-MES on constrained problems from 6 to 100 dimensions, with MF-MES impractical beyond roughly 8 dimensions.
- The predicted variant outperforms the evaluated one when high-fidelity predictions are reliable (Hartmann, Borehole, Ackley) and the reverse when they are not (Rosenbrock, solar7).
Strengths
Addresses a genuine methodological gap
Secs. 1 and 3 target four properties jointly — high dimension, black-box constraints, arbitrary fidelity count, non-nested designs — and deliver a concrete algorithm (Alg. 2) with an open-source BoTorch implementation (github.com/Nemo-Virtual-Lab/MF-SCBO).
Elegant, zero-overhead LOO safeguard
The closed-form leave-one-out derivation for the sequential autoregressive GP (Eqs. 9–14, criterion Eq. 15) protects the predicted trust-region center without refitting any GP; this is the paper's most original technical component.
Broad and honest empirical evaluation
Candid reporting of failures
Weaknesses
Single, weakened multi-fidelity baseline
No theoretical analysis
Statistical reporting too weak for the claimed rankings
Headline claim is contradicted in parts by the paper's own tables
Restrictive assumptions and underspecified details
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
MF-SCBO: Structured Preliminary Review
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 4 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 MF-SCBO?
MF-SCBO is a multi-fidelity extension of the SCBO constrained Bayesian optimization algorithm. It combines Gaussian-process surrogates across several fidelity levels with trust regions and Thompson sampling, and adds a leave-one-out check that keeps the trust-region center reliable when it is based on predictions rather than actual high-fidelity evaluations. It targets problems that are high-dimensional (up to 100 variables in the paper) and subject to black-box constraints.
Does MF-SCBO beat single-fidelity SCBO?
Usually in convergence speed on the tested 6–100 dimensional constrained benchmarks, but not always: on Rosenbrock40D/100D, single-fidelity SCBO reached better final objective values within the same 4000-evaluation budget, and on problems with unreliable high-fidelity predictions (solar7) the safer 'evaluated' variant or SCBO can be preferable.
Is code available for MF-SCBO?
Yes. The authors released a Python implementation built on the BoTorch library at github.com/Nemo-Virtual-Lab/MF-SCBO (linked in Section 4 of the paper).
What are the main limitations of MF-SCBO?
There are no theoretical convergence guarantees; constraints must be independent of the fidelity level and their evaluation costs are not modeled; the only multi-fidelity baseline compared was a penalized, time-limited MF-MES; and results are reported as means ± 1.96 standard deviations without significance testing.
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.
- Convergence figures (Figs. 3, 4, 8–10) are the central evidence but cannot be analyzed from the extracted text; this review relies on Tables 1–2 and the prose descriptions.
- Appendix B ablation results are reported only as figures, so their magnitudes cannot be checked.
- Minor extraction/typo artifacts exist (e.g., 'MFMEF' in Table 1, 'height-dimensional' for Borehole, sum-index typos in Eq. 6).
- The TS candidate-set size r and some hyperparameters are not stated in the text; reproducibility assessment assumes the released code fills these gaps.
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.