An Adaptive Competitive Surrogate Modeling Framework for Efficient Control Parameter Optimization of Buck Converters

Controller parameter tuning for buck converters is computationally expensive because it requires repeated high-fidelity simulations. Moreover, the prediction performance of individual surrogate models may vary across different regions and stages of the optimization process, whereas fixed multi-surrogate combinations cannot readily adapt to such changes. To address these issues, this paper proposes a competitive surrogate selection-assisted genetic algorithm (CSS-GA) for efficient controller parameter optimization of a buck converter. The proposed framework employs a heterogeneous surrogate pool consisting of Kriging, Support Vector Regression (SVR), and a Radial Basis Function (RBF) model. A competitive selection mechanism dynamically evaluates the surrogate models according to their cross-validated prediction accuracy and uncertainty information and selects a single surrogate for evolutionary fitness prediction, thereby avoiding reliance on a fixed weighted ensemble for population fitness evaluation. In addition, an Expected Improvement (EI)-enhanced infill mechanism dynamically combines surrogate information for acquisition and selects informative candidates for subsequent high-fidelity evaluation and model retraining. Simulation results show that CSS-GA reduces the required high-fidelity simulation evaluations by 87.5% under the adopted experimental protocol while maintaining effective optimization performance. Comparisons with standard genetic algorithms (GAs), single-surrogate and ensemble-based variants, and representative dynamic surrogate-assisted evolutionary algorithms—including ASMEA, SAEA-HAS, AS-SMEA, and HSSM—further evaluate the performance of the proposed framework. Under the adopted comparison protocol, CSS-GA achieves the lowest mean final fitness among the compared dynamic surrogate-assisted methods. The results demonstrate the effectiveness of the proposed framework for reducing the high-fidelity evaluation cost of controller parameter optimization in the investigated buck-converter problem.