Initialises the global net structure for the OGP toolbox.
ogp(nin,nout,covarfn,covpar, prmean, prmeanp)
ogp(nin, nout, covarfn,covpar) creates a Gaussian Process model in the
GLOBAL structure net which has nin input and nout output
dimensions.
The string in the field covarfn specifies the type of the
covariance function to be used. The parameters to the covariance function
are given in covpar. The available covariance functions are listed in
ogpcovarf.
The function returns a data structure net with the rest of
parameters set to zero. If covpar is not specified then the
default values are assigned to it (description of implemented
kernel functions is in ogpcovarf).
ogp(nin, nout, covarfn, covpar, prmean, prmeanp) also sets a
prior mean function to the Gaussian process. The address in prmean is
the function returning the prior means at locations currX. The field
prmeanp contains optional parameters to prmean.
The function prmean has the structure
[meanVec] = prmean(x,prmeanp);
The likelihood function and its parameters can be set in ogpinit,
similarly to prmean (see ogpinit).
Additional parameters which influence the computation of the posterior GP
and the re-calculation of the hyperparameters are changed directly, using
the designed fields of the structure net.
The structure net is global, thus there is no need for it to be
transmitted as a parameter.
The fields in the structure net are:
type - fixed, the string 'ogp';
nin,nout- dimension of inputs and outputs respectively;
thresh - lower limit to the score of a BV (see ogpbvmin),
no input with smaller score is included;
maxBV- maximal size of BV set.
covarfn,kpar - parameters specifying the covariance function
(see ogpcovarf);
bias - logarithm of constant offset in the covariance function
(see ogpcovarp);
inweights- logarithm of inverse length scale for each input;
prmean,prmeanp - address to the prior mean function and field
containing additional parameters. If prmean is unspecified, then the
prior mean is zero.
BV - elements of the Basis Vector (BV) set (this
matrix is initially empty);
w,C- parameters of the posterior mean and variance (initially empty);
isBVfixed - whether to modify or not the set of Basis
Vectors BV. Keeping it fixed speeds up the inference process, buth
then one has to set the BV set in advance (using e.g. ogpemptybv or
ogptrain with a small subset of the data and isBVfixed set
to zero).
proj - binary indicator variable specifying the KL-divergence
to use. The default is the KL-divergence that matches the moments
(nonzero value to proj) and the negation leads to matching the
mean and variance at the BV locations.
outtype - the type of the ``training'' data; a boolean-like
variable that can be either 'direct', which means that yTrain
must be of dimensionality nout. The string can also be
'indirect', meaning that there is no need for the dimension of
yTrain to match net.nout but there has to be an ``desired
output''. The third choice is 'none', meaning that there is no
desired output -- this is the case of unsupervised learning e.g. in
density estimation.
KB,KBinv - storing the kernel matrix of the BV set and
its inverse - to speed up calculations.
hyplambda,hypmean- values of prior precision and mean of the OGP hyperparameters. This is often needed to ensure numerical stability of the code.
likpar - the vector of the likelihood parameters (see
ogpinit), specific to each likelihood function.
likaddr,likoptfn - the address of tghe function to recompute
the likelihood parameters and the vector of optional parameters
used in the calculation (see e.g. em_gauss or em_lapl).
ogpinit, ogpcovarf, ogppak, ogpunpak, ogpfwd
Copyright (c) Lehel Csató (2001-2004)