Ensemble Surrogate-Assisted Optimization with Dynamic Dimensional Splitting for High-Dimensional Expensive Mixed-Variable Problems

Abstract

High-dimensional expensive mixed-variable optimization presents a significant challenge due to the curse of dimensionality and variable heterogeneity. Existing methods often struggle with blind dimension reduction and inefficient search guidance. This paper proposes a novel framework: Ensemble Surrogate-assisted hybrid search with importance-aware Dimensional Splitting (ESDS). Unlike traditional blind splitting, ESDS dynamically allocates the dimensional budget to the most critical variables via data-driven sensitivity analysis. To balance global exploration and local exploitation, a hybrid search mechanism is introduced to the predicted best individual, which is generated based on an ensemble surrogate. It competes with the historical best individual to guide the initialization of the trust region-based Bayesian optimization. Furthermore, an adaptive heterogeneous ensemble is employed, aggregating dynamic weighted predictions to mitigate single-model fragility. Experiments on seven benchmarks against six state-of-the-art baselines demonstrate ESDS's superior convergence efficiency.