mgpr: An R package for multivariate Gaussian process regression

Abstract

Gaussian process regression (GPR) is a non-parametric kernel-based machine learning method. GPR is based on Bayesian formalism, which enables the estimation of prediction uncertainty of the response variables. We propose an R package that provides an easy-to-use interface for multivariate GPR. The mgpr package was originally developed for remote sensing-based forest inventories that require multivariate prediction of forest attributes. The mgpr package supports both univariate and multivariate responses using a separable kernel and includes a robust hyperparameter estimation algorithm. The mgpr package is suitable for various regression problems with single response or multiple responses and provides good prediction performance.

Keywords

• Bayesian approach
• Covariance function
• Stochastic process