Title: Unifying Surrogate and Criterion Control in SAEA via Meta-Black-Box Optimization for Enhanced Robustness and Accuracy

Surrogate-assisted evolutionary algorithms (SAEAs) have become a standard approach for tackling expensive black-box optimization challenges. However, their performance is often constrained by rigid, hand-crafted components that hinder adaptability and generalization across diverse tasks. While Meta-black-box optimization (MetaBBO) offers a promising avenue for the adaptive configuration of algorithmic parts, current methods typically manage only isolated components. There is a notable lack of research on the unified control of multi-component systems like SAEAs within this framework. Furthermore, the critical trade-off between robustness and accuracy in surrogate modeling—essential for balancing early-stage exploration with late-stage exploitation—has been largely overlooked in existing literature.

To bridge these gaps, we introduce AdaE-SAEA, an adaptive ensemble surrogate-assisted evolutionary algorithm designed for expensive multi-objective optimization. This method integrates SAEA as the low-level optimizer within a MetaBBO framework, simultaneously governing both the infill criterion and ensemble-based surrogate modeling. The architecture employs bagging and boosting modules to dynamically adjust the balance between robustness and accuracy according to the current search phase. Concurrently, a meta-policy drives adaptive sampling by selecting the appropriate infill criterion. This meta-policy is optimized via reinforcement learning, utilizing parallel sampling and centralized training to enhance both computational efficiency and transferability.

Empirical evaluations on both synthetic and real-world benchmarks indicate that AdaE-SAEA surpasses state-of-the-art baselines and other MetaBBO-based approaches. Additionally, the study confirms the efficacy of using TabPFN as a foundational surrogate model for ensemble learning. To our knowledge, this research represents the first effort to consolidate the control of surrogate modeling and infill criteria within SAEAs while explicitly managing the robustness-accuracy trade-off.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC