Calculation of the sparse posterior
ogptrain(x_train, y_train)
ogppost(x_train,y_train,its) - considers the GLOBAL (prior) Gaussian
Process structure net together with the training data
[x_train,y_train] and sets the (posterior) Gaussian Process structure
net using the training data within the online learning algorithm. The
number of sweeps through the dataset is its, if unspecified, its
default value is 1.
In Sparse online learning, by default the BV set changes, which
implies significant matrix manipulation. If we want the BV set to be
constant, we need to set the indicator isBVfixed (e.g.
net.isBVfixed=1). This makes the code faster, but it also implies
that one has to set the BV set ``manually'' (using ogpemptybv).
Alternatively one should first make an iteration using the default BV
set selection mechanism and then set net.isBVfixed=1.
The functions ogpost and ogptrain use other two GLOBAL
structures ogptrain and ep. Setting the fields of these
structures influences the way the posterior is computed.
ogppost(x_train,y_train) - uses the GLOBAL structure
gpopt to transmit/receive additional variables from/to the training,
like eg. the TAP/EP extension of the online learning, the test- or training
errors and the various options to perform hyperparameter optimisation (see
defoptions for details).
Two sets of hyperparameters are distinguished: those related to the
kernel function and those characterising the likelihood. The initialisation
of the kernel parameters is done in ogp and those of the likelihood
function in ogpinit (see the respective functions).
Optimisation of the hyperparameters is implemented within the function
ogptrain.
ogp, ogpinit, ogptrain, ogpstep_full, ogpstep_sp, defoptions
Copyright (c) Lehel Csató (2001-2004)