Projection pursuit Gaussian process regression

A primary goal of computer experiments is to reconstruct the function given by code via scattered evaluations. Traditional isotropic Gaussian process models suffer from curse dimensionality, when input dimension relatively high limited data points. with additive correlation functions are scalable but they more restrictive as only work for functions. In this work, we consider a projection pursuit model, in which nonparametric part driven an regression. We choose higher than original dimension, and call strategy “dimension expansion”. show that expansion can help approximate complex gradient descent algorithm proposed model training based on maximum likelihood estimation. Simulation studies method outperforms traditional models.