Marginalise over the hyperparameters

[landmanbester]

The current algorithm trains the Gaussian process on (possibly mock) input data and then simply uses the best fitting values of the hyperparameters throughout. This is not really correct from a Bayesian point of view. The hyperparameters can be marginalised by using a sequential Monte-Carlo sampler, resulting in a so called Metropolis within Gibbs sampling scheme (see for example here). This will be required for realistic tests of the CP.