The GP-Huber model proposes Gaussian process with likelihood of data expressed as Huber density function. The hybrid Monte Carlo and Laplace approximation methods are proposed to solve the resulting complex inference. The proposed model is demonstrated to provide good trade off between efficiency and robustness for the cases of error following (i) normal, (ii) Student-t, (iii) Laplace, and (iv) Cauchy distribution. We also show performance of the model when vertical outliers and bad leverage points are added to the training data.
Please add gpstuff package in the path to run the scripts in the Experiment section which can be found at https://github.com/gpstuff-dev/gpstuff.