Computes the gradient for the Sparse OGP.
covf = ogpevidgrad(hp);
g = ogpevidgrad(hp)
takes the OGP parameter vector hp
and the
GLOBAL Sparse OGP data structure net
and returns the gradient of the
marginal likelihood (i.e. the gradient of the evidence).
The gradient is computed only with respect to the kernel parameters, the
gradients of the evidence w.r.t. the likelihood parameters are treated
differently. g
is a column vector with the dimension of model
hyperparameters where the order is provided by ogppak
.
This function uses ogpcovgrad
to compute the gradient values. If the
structure ep
is empty, then the gradient (and the evidendce) is only
approximate, this might lead to inaccuracies when optimising the kernel
parameters.
The computation of the gradient is can be extremely time and
memory-consuming. To alleviate the memory requirements, if there are too
many elements to compute, ogpevidgrad
makes a sequence of gradient
computations, i.e. calls ogpcovgrad
with an optional argument -- see
the documentation of ogpcovgrad
for details.
ogpcovarf
, ogppak
, ogpcovgrad
, ogptrain
Copyright (c) Lehel Csató (2001-2004)