Gaussian Process Optimization Package in R

Many Gaussian process packages are available in R. For example there is $\textbf{BACCO}$ that offers some calibration techniques, $\textbf{mlegp}$ and $\textbf{tgp}$ focusing on treed models and parameter estimation and $\textbf{GPML}$ for Gaussian process classification and regression.

The problem with these packages is that the choice of correlation function is restricted. Only some choices are provided for building the correlation function (Gaussian, Matern, etc...).

Does anyone have specific experience on how to can insert my own correlation function and just use the optimization routines available in these packages which are specifically tailored for GP's. Or is there a package which allows me to do that ?

• I feel your pain. I went down this path a while back and got about 70% of the way to releasing a package that was more robust and customizable ... then I changed jobs and started school again and never really wrapped it up. Basically, the way I cracked it was to piggy back on stan to integrate out kernel hyperparameters and expose the stan program to the user. – Sycorax Aug 26 '16 at 23:16
• @GeneralAbrial thanks for your comment. Based on your experience, what optimizer best fits maximizing the multivariate normal likelihood specifically in R. – Wis Aug 26 '16 at 23:18
• The problem isn't about optimizing a mvn density at all -- it's tuning hyperparameters. So I punted and used stan to integrate them out. – Sycorax Aug 26 '16 at 23:21
• @GeneralAbrial typically in the GP we aim at finding the hyperparatmeter estimates through an MLE approach or a Bayesian posterior MCMC approach. You mean by tuning is finding a good starting point for your hyperparameters ? – Wis Aug 26 '16 at 23:23
• No, I'm saying stan is how I sampled ("tuned") hyperparameters. – Sycorax Aug 26 '16 at 23:25