Hybrid-Surrogate-Assisted Particle Swarm Optimization with Adaptive Surrogate Selection for Solving High-Dimensional Expensive Optimization Problems

This paper proposes a hybrid-surrogate-assisted particle swarm optimization (HSAPSO) for high-dimensional expensive optimization problems (EOP) being inherently complex and computationally costly. To address the challenge that a single surrogate model may struggle to provide efficient exploration through the entire evolutionary process, HSAPSO develops strategies consisting of adaptive surrogate selection and hyperparameter estimation. Initially, a linear surrogate facilitates exploration with limited data, later switching to a radial basis function (RBF) surrogate for nonlinear approximation as data grows. Maximum likelihood estimation (MLE) fine-tunes RBF hyperparameters ensuring high accuracy and robustness. Social learning particle swarm optimization with progressive sampling balances exploration and exploitation. Benchmark tests show HSAPSO outperforms four state-of-the-art, demonstrating its effectiveness for high-dimensional expensive optimization.