Computes the KL-distance of the GP marginals.
kl = ogpkl(net2, sameBV);
[kl]=ogpadjgp(net2) computes the KL-distances between two
Gaussian processes which use the same covariance function. One of the
Gaussian processes is given in the GLOBAL structure net, the second is
ginven as a parameter.
The same kernel function for the two processes is essential: this is needed to compute the KL-distance between GPs (exact computation is only possible in this case).
The computation is much faster if the BV sets for the GLOBAL net and
net2 are the same (i.e. when using a GP inference with fixed BV set or
transductive learning). This is indicated with a third boolean argument
sameBV set to a nonzero value.
This function can be used as a Cauchy-termination criterion to the online iterations, if the successive OGP structures are close in the KL-sense, this means that the aflgorithm has reached a stable solution.
ogp, ogptrain, ogpcovarf, ogpcovarp
Copyright (c) Lehel Csató (2001-2004)