Two-stage attention-guided active learning for nonstationary Gaussian process modeling in robust parameter design
Robust parameter design is vital for continuous quality improvement in industrial practice. However, modeling nonstationary response surfaces with limited data remains challenging because insufficient predictive accuracy may introduce optimistic bias in estimating optimal parameter settings. We propose a two-stage attention-inspired active learning method within a nonstationary Gaussian process (NSGP) modeling framework for robust parameter design. This two-stage acquisition first ranks candidates by expected quality loss and then applies a maximin rule to preserve subset diversity. In the first stage, candidate design points are ranked by their relevance to the specified target region, and a subset with high relevance scores is selected for further refinement. In the second stage, spatial diversity is enforced by selecting additional points from this subset to mitigate clustering. The selected points are used to update the training set and NSGP model. A genetic algorithm solves the NSGP-based robust parameter design problem and determines optimal settings. Comparisons with classical design baselines and three modern acquisition criteria show that the proposed method improves predictive accuracy and enhances the robustness of solutions under limited data. The effectiveness of the method is further supported by ablation experiments, target-value sensitivity analysis, and a modified 10-variable Wing Weight analytical benchmark.
Authors
- Weidong Huang
- Zebiao Feng
- Jianjun Wang
- Xu Yang
Publication Details
- Journal
- Figshare
- Published
- 2026-11-20
- DOI
- https://doi.org/10.6084/m9.figshare.35074020
- Primary Topic
- Optimal Experimental Design Methods
- Type
- article
- Field-Weighted Citation Impact
- 0.00