A Dynamic Aggregation Strategy Enhanced Efficient Global Optimization Algorithm for Solving High-Dimensional Turbomachinery Design Problems
Qineng Wang, Zhendong Guo, Yun Chen, Guangjian Ma, Liming Song, Jun Li
cs.LG · cs.CE
Quick verdict
DA-EGO dynamically regroups subspaces for high-dimensional expensive optimization; benchmark and CFD-case evidence is broadly supportive despite minor flaws.
Summary
The paper proposes DA-EGO, a surrogate-based global optimizer for high-dimensional (30–90 variables) expensive black-box problems such as CFD-driven turbomachinery design. It decomposes the design space into low-dimensional subspaces, runs EGO (kriging plus expected improvement) within each, and merges subspace optima into an elite point. Grouping is dynamic rather than fixed: polynomial-chaos Sobol sensitivity ranks suspected interacting variable pairs, and a surrogate-based perturbation analysis confirms or rejects them, with confirmed pairs accumulating across cycles; a space-reduction step shrinks each subsearch to the best-performing half of 10,000 surrogate samples. On 21 benchmark instances (seven shifted/rotated CEC-style functions at 30, 60, and 90 dimensions; budget 1,500 evaluations; 20 runs), DA-EGO attains the lowest mean on 17 instances against GA, DE, GSK, IKAEA, GSGA, Nash-EGO, and a random-grouping ablation (RG-EGO), while conceding that GSGA wins on shifted Rosenbrock at 60 and 90 dimensions. Wilcoxon signed-rank tests support most comparisons. Two CFD cases—28-variable NASA Rotor 37 (+1.65% efficiency in 1,000 evaluations) and a 60-variable two-stage compressor (+1.07%)—demonstrate practical value. Main limitations: an arithmetically impossible standard deviation in Table 3 (F7, 30-D), Nash-EGO silently absent from the 60/90-D tables, an unspecified grouping-size parameter, no structure-exploiting Bayesian-optimization baselines, and 35–40-minute runtimes that confine the method to genuinely expensive objectives.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 2
● low 2
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
- Dynamic subspace decomposition with per-cycle interaction-based regrouping lets EGO-scale surrogates solve 30–90-dimensional expensive black-box problems within roughly 1,500 function evaluations.
- Subspace-level interaction mining (PCE-Sobol ranking plus perturbation confirmation) reduces the sampling cost of variable-interaction detection enough to make decomposition viable for expensive problems.
- Across 21 benchmark instances, DA-EGO achieves the lowest mean on 17, dominating separable and partially separable functions while losing to GSGA on non-separable shifted Rosenbrock at 60/90-D.
- In two CFD case studies (28-variable Rotor 37; 60-variable two-stage compressor), DA-EGO yields the largest efficiency gains (+1.65% and +1.07%) under a 1,000-evaluation budget.
Strengths
Unusually honest negative reporting
The abstract and Section 4.3 explicitly concede GSGA wins on shifted Rosenbrock (F7) at 60/90-D and mark several comparisons ≈; Table 6 discloses the full cost side (DA-EGO 35–40 min/run vs 0.124–3.2 s for DE/GA/GSK and 68–149 s for IKAEA) instead of hiding algorithmic overhead.
Ablation isolates the core contribution
RG-EGO (random grouping, Section 4.1) controls for the interaction-analysis machinery: on partially separable F4 at 30-D the mean gap is 7.672E+04 (DA-EGO) vs 1.285E+06 (RG-EGO) and 2.387E+05 (IKAEA), directly evidencing that learned grouping — not decomposition per se — drives the gain.
Cost-aware redesign of variable-interaction detection
Realistic engineering validation
Sound experimental protocol
Weaknesses
Arithmetically impossible statistic in Table 3
Baseline set silently shrinks at higher dimensions
Incomplete algorithmic specification
Ambiguous ablation composition
No structure-exploiting BO baselines and weak engineering statistics
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
DA-EGO targets high-dimensional (30–100 variables) expensive black-box problems where each objective evaluation costs hours of CFD. The claimed contributions are: (i) decomposition of the design space into low-dimensional subspaces solved by EGO (kriging + expected improvement), with subspace optima aggregated into an elite point; (ii) a dynamic decomposition in which variable groupings are rebuilt each cycle from interaction knowledge mined in subspaces — PCE-based Sobol sensitivity (Section 2.3) ranks suspected interacting pairs, and surrogate-based perturbation analysis (Eq. 14, threshold H_ij ≥ 0.2) confirms or rejects them; and (iii) adaptive per-subspace search-range reduction (Algorithm 2: keep the best-performing half of 10,000 surrogate samples). Evidence comprises 21 benchmark instances (7 CEC-style functions at 30/60/90-D, 1,500-NFE budget, 20 runs) plus two CFD cases (28-variable Rotor 37; 60-variable two-stage compressor). The abstract and Section 4.3 are notably candid: GSGA beats DA-EGO on shifted Rosenbrock at 60/90-D, and several Wilcoxon comparisons are marked ≈.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 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 DA-EGO and what problem does it solve?
DA-EGO (Dynamic Aggregation EGO) is an algorithm for high-dimensional expensive black-box optimization — problems with roughly 30–100 design variables where each objective evaluation (e.g., a CFD simulation) takes hours. It splits the design space into low-dimensional subspaces, optimizes each with EGO (kriging surrogate plus expected improvement), and rebuilds the variable groupings every cycle using learned variable-interaction information, so knowledge from cheap subspace searches accumulates into global search guidance.
How well does DA-EGO perform compared with other algorithms?
On 21 benchmark instances (seven functions at 30, 60, and 90 dimensions, 1,500-evaluation budget, 20 runs each), DA-EGO achieved the lowest mean result on 17 instances against GA, DE, GSK, IKAEA, GSGA, Nash-EGO, and a random-grouping variant. However, GSGA was better on the non-separable shifted Rosenbrock function at 60 and 90 dimensions, so the advantage depends on the problem's interaction structure. In CFD case studies it improved Rotor 37 efficiency by 1.65% and a two-stage compressor's by 1.07% within 1,000 evaluations.
Is code or data available for DA-EGO?
The paper's data-availability statement points to a public repository (github.com/zhet1997/DA_EGO_publish). The benchmark functions are standard CEC-style large-scale functions, so numerical results are plausibly reproducible; the engineering cases require the commercial NUMECA CFD suite and cannot be re-run independently. Some algorithm parameters (e.g., a group-size cap in the pseudocode) are not specified in the text.
What are the main limitations of DA-EGO?
Three stand out: (1) cost — surrogate construction and interaction analysis make each benchmark run take 35–40 minutes versus under a second for DE or GSK, so the method only pays off when each true evaluation is expensive; (2) scope — it wins on separable and partially separable problems but loses to GSGA on the non-separable shifted Rosenbrock at 60/90 dimensions; (3) reporting — reviewers noted an impossible standard-deviation value in one results table and a comparison algorithm missing from the higher-dimension tables.
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-curve figures (Figs. 4–8) are not analyzable from the extracted text; performance judgments rest on tables and prose.
- The extraction contains duplicated near-identical passages (Sections 4.4 and 5 each appear twice with different figure numbering), so some section/figure inconsistencies may be extraction artifacts rather than manuscript errors.
- Several equations are garbled by extraction (kriging MSE in Eq. 3, the Rosenbrock exponent in Table 2); formula interpretation relies on standard forms.
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.