Nested Latin Hypercube-Based Sampling for Efficient Uncertainty Quantification Using Sensitivity-Assisted Least Squares SVM

Abstract

Recently, a methodology to use the sensitivity information for building the least squares support vector machine (LS-SVM)-based surrogate model for uncertainty quantification in the context of circuit systems was proposed. It was shown that the sensitivity-enhanced LS-SVM could successfully reduce the simulation data required for building LS-SVM-based surrogate models. However, the number of samples required for building the surrogate models is not known a priori. In this article, we present an iterative technique that uses the nested Latin hypercubes to add the samples until the surrogate model achieves the desired accuracy. The presented technique is demonstrated using two numerical examples, where we show that the proposed method can significantly reduce the amount of simulation data required for building LS-SVM-based surrogate models.