Research on Applicability and Efficient Construction with Typical Surrogate Models

Abstract

Surrogate models (SMs) are generally employed in expensive engineering optimization and uncertainty quantification. At present, it is still one of the hot topics for researchers to construct high-performance SMs with limited resources. By comparing the predicted errors of the typical functions, this work intends to obtain the applicability of typical SMs and the count of the necessary experimental samples for efficiently constructing SMs. In detail, representative test functions such as Branin-Hoo, Camelback, Godlstein-Price, Hartman, ZDT3 and ZDT4 are approximated by the models such as PRS, RBF, GP, NN and SVR. The predicted indicators such as R2, RMSE and MAE under multiple groups of experimental samples are discussed by cross-validation. Finally, the applicability of typical SMs and the count of experimental samples in DOE are given, and the conclusion might be a reference for constructing high-performance SMs and an ensemble of SMs.