Uncertainty-Adaptive Robust Surrogate Aggregation for Expensive Optimization Problems with Noisy Historical Data
Year: 2026
Type: conference-paper
Abstract: Offline data-driven evolutionary algorithms (DDEAs) have become an effective paradigm for solving expensive optimization problems in autonomous unmanned systems under limited historical data. However, most existing DDEAs aggregate surrogate predictions using arithmetic averaging, making prediction quality highly dependent on the reliability of historical data. When historical data are partly corrupted, inconsistent surrogate predictions may be introduced, and the reliability of ensemble-based optimization decreases. Motivated by this issue, we develop a Mean with Adaptive Trimming (MAT) strategy as an adaptive surrogate aggregation mechanism inspired by robust statistics. Unlike conventional robust aggregation schemes, MAT adaptively adjusts the aggregation process based on prediction uncertainty. The final prediction is then obtained by mean aggregation over the adaptively trimmed prediction set. This formulation enhances robustness to extreme predictions while preserving useful information from the surrogate ensemble. MAT is integrated into two representative DDEAs and evaluated on five benchmark functions under six noise conditions, including Gaussian, Cauchy, and mixed noise. Across these functions and noise scenarios, MAT remains competitive under noise-free and Gaussian noise settings while achieving clear improvements under heavy-tailed and mixed-noise settings. Ablation studies further demonstrate that adaptive trimming combined with mean aggregation is more suitable for surrogate ensembles than conventional robust aggregation strategies.
Authors: Zi-Lu Su, Tao Jin, Lin Shi, Jian-Yu Li
Language: English
Subfield: Computational Theory and Mathematics
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204435131