Grammar based Kernel Composition Design using Gaussian Processes for Non-Parametric Regression

Abstract

This Thesis describes a method to automatically compose a kernel to use in a non-parametric causal inference design. Choosing a kernel that provides a good fit for the model in Gaussian Process regression can be challenging. Through summation and multiplication composite kernels get created, by comparing and elaborating the composite kernels results in the best fitting kernel. This method is an extension to the Bayesian non-parametric quasi-experimental design.