Comparative Study of Kriging and Support Vector Regression for Structural Engineering Applications

Abstract

Metamodeling techniques have been widely used as substitutes for high-fidelity and time-consuming models in various engineering applications. Examples include polynomial chaos expansions, neural networks, kriging, and support vector regression (SVR). This paper attempts to compare the latter two in different case studies so as to assess their relative efficiency on simulation-based analyses. Similarities are drawn between these two metamodel types, leading to the use of anisotropy for SVR. Such a feature is not commonly used in the SVR-related literature. Special care was given to a proper automatic calibration of the model hyperparameters by using an efficient global search algorithm, namely the covariance matrix adaptation–evolution scheme. Variants of these two metamodels, associated with various kernel and autocorrelation functions, were first compared on analytical functions and then on finite element–based models. From the comprehensive comparison, it was concluded that anisotropy in the two metamodels clearly improves their accuracy. In general, anisotropic L2-SVR with the Matérn kernels was shown to be the most effective metamodel.