Reliability assessment by adaptive kernel-based surrogate models - Approximation of non-smooth limit-state functions

Abstract

Adaptive surrogate models are of practical use for reliability analysis based on costly-to-evaluate limit-state functions. The quality of the approximation made depends on both the selected type of surrogate model (and its related assumptions) and the adaptive scheme applied for the construction of the approximate model. Most of surrogate models assume some degree of smoothness, which allows them to be learned with a not so large set of input-output data pairs. This paper investigates the use of Matérn kernels in the context of support vector regression, with tuned regularity parameters. This kernel is used in an adaptive scheme based on MCMC sampling, whose objective is to progressively sample the failure domain. The proposed approach is applied to both a smooth and a non smooth limit-state functions, showing the benefits of using such a highly flexible kernel.