Title: Establishing Confidence in Black-Box Optimization: An All-Encompassing Approach to Explainability

Abstract: The optimization of expensive black-box functions under strict evaluation limits poses substantial hurdles across numerous practical scenarios. While Surrogate Optimization (SO) is a prevalent solution, its inherent opacity—stemming from complex surrogate models and sampling mechanisms like acquisition functions—frequently results in poor transparency and a lack of interpretability. Although current research has largely focused on improving convergence toward global optima, the practical interpretability of novel strategies remains insufficiently investigated, particularly within batch evaluation contexts. To address this, we introduce Inclusive Explainability Metrics for Surrogate Optimization (IEMSO), a robust collection of model-agnostic metrics aimed at bolstering the transparency, reliability, and explainability of SO methodologies. These metrics offer both intermediate and post-hoc insights, enabling practitioners to build trust before and after conducting costly evaluations. We organize these metrics into four main categories, each addressing a distinct facet of the SO workflow: Sampling Core Metrics, Batch Properties Metrics, Optimization Process Metrics, and Feature Importance. Our empirical results highlight the considerable promise of these proposed metrics across various benchmark datasets.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC