View article

Kernel based surrogate models like Kriging are a popular remedy for costly objective function evaluations in optimization. Often, kernels are required to be definite. Highly customized kernels, or kernels for combinatorial representations, may be indefinite. This study investigates this issue in the context of Kriging. It is shown that approaches from the field of Support Vector Machines are useful starting points, but require further modifications to work with Kriging. This study compares a broad selection of methods for dealing with indefinite kernels in Kriging and Kriging-based Efficient Global Optimization, including spectrum transformation, feature embedding and computation of the nearest definite matrix. Model quality and optimization performance are tested. The standard, without explicitly correcting indefinite matrices, yields functional results, which are further improved by spectrum transformations.