Inference for Sparse Gaussian Processes.
ogptrain(x_train, y_train)
The function ogptrain(x_train,y_train) takes the GLOBAL Gaussian
Process data structure net and performs the following operations:
approximates the posterior process.
adjusts covariance function parameters.
adjusts likelihood function parameters.
The above steps constitute a single cycle in the EM algorithm built for the joint optimisation of the posterior process and the hyperparameters.
Notice that - due to the changes in the hyperparameters - the posterior
process at the end of ogptrain is no longer optimal. Thus, if
prediction is wanted, then one should perform an extra computation step
using ogppost. This is not done in ogptrain.
The calculations are governed by the fields of the GLOBAL structure
gpopt: the indicators influencing calculation of the posterior are
grouped into the sub-structure gpopt.postopt.
Given an approximation to the posterior process (step 1), we then adjust
the parameters of the covariance kernel and the likelihood. The covariance
parameters are optimised using a conjugate gradient algorithm, the specific
algorithm and the number of steps can be altered via the structure
gpopt.covopt.
The optimisation of the likelihood parameters is done as the last step of
the EM procedure. It is independent of the optimisation of covariance
function parameters. It can be done using gradients (like g_l_gauss)
or EM (see em_gauss).
ogp, ogppost, ogppak, ogpevid, ogpevidgrad
Copyright (c) Lehel Csató (2001-2004)