GP-MC:  Use Markov chain to sample Gaussian process hyperparameters.

The gp-mc program is the specialization of xxx-mc to the task of
sampling from the posterior distribution of the hyperparameters of a
Gaussian process model, and possibly also from the posterior
distribution of the latent values and/or noise variances for training
cases.  If no training data is specified, the prior will be sampled
instead.  See xxx-mc.doc for the generic features of this program.

The primary state of the simulation consists of the values the
hyperparameters defining the Gaussian process. These hyperparameters
are represented internally in logarithmic form, but are always printed
in 'sigma' form (ie, with in the units of a standard deviation).

This state may be augmented by the latent values for training cases
and/or by the noise variances to use for each training case.  This
auxiliary state is updated only by the application-specific sampling
procedures described below.  This augmentation is not needed for
regression models where the noise variance is the same for all cases,
but is needed for classification models, models for count data, and
regression models with varying noise (equivalently, noise that has a t
distribution).

If latent values are present when they are not needed, they will be
used when updating the hyperparameters (treated as noise-free data).
This will probably slow convergence considerably, but might be desired
in order that these values can be plotted.  The best way to proceed in
this situation is to generate the values at the end of each iteration
but discard them before sampling resumes.  One MUST discard the latent
values before updating hyperparameters when using a regression model
with autocorrelated noise.

The following application-specific sampling procedures are implemented:

   scan-values [ N ]

       This procedure does N Gibbs sampling scans for the latent values
       associated with training cases, given the current hyperparameters 
       (and possible case-specific noise variances).  If N is not 
       specified, one scan is done.  Note that a single scan is 
       generally not sufficient to obtain a set of values that are
       independent of the previous set.

       If latent values do not exist beforehand, this procedure operates 
       as if the previous values were all zero.  

       This operation is not allowed for regression models with
       autocorrelated noise.

   met-values [ scale [ N ] ]

       Does N Metropolis updates of the latent values (default is one),
       using a Gaussian proposal distribution whose mean is the current
       set of latent variables and whose covariance matrix is the prior
       covariance of the latent variables multiplied by scale^2 (with
       default of one).  

       As a special fudge (probably useful only for testing), a negative 
       scale can be given, in which case the covariance matrix for the
       proposal is diagonal, based simply on the prior variances.

       If latent values do not exist beforehand, this procedure operates 
       as if the previous values were all zero.  

   mh-values [ scale [ N ] ]

       Does N Metropolis-Hastings updates of the latent values (default 
       is one), in which the proposed state is found by multiplying all
       the latent values by sqrt(1-scale^2) and then adding Gaussian noise 
       with mean zero and covariance given by the prior covariance of the 
       latent values multiplied by scale^2.  The default for scale is one, 
       in which case the proposal is a sample from the prior, but this is 
       probably not good for most problems.  

       If latent values do not exist beforehand, this procedure operates 
       as if the previous values were all zero.  

   init-values

       Initialize latent values associated with training cases by setting
       them to values determined by the target(s) associated with that
       case and by the distribution of values for all training cases.
       Details vary with the model; some models may not implement this
       operation.  This operation is ignored if latent values already exist.

   sample-values [ old ]

       Samples latent values associated with training cases given the 
       current values of the hyperparameters, ignoring the current latent
       values (if any).  

       This operation is possible only for regression models.  It is 
       allowed for regression models with autocorrelated noise, but as
       noted above, in this context, they will probably have to be 
       discarded before doing anything else (and hence are useful only 
       in allowing quantities such as "b" to be evaluated).

       If the optional "old" argument is present and is non-zero, an 
       older, slower, version is used.  This version is also used when
       the model has autocorrelated noise.  The two versions sample from
       the same distribution, but don't produce the same values with the
       same seed.

   discard-values

       Eliminates the latent values currently stored.  This is useful
       for regression models, where keeping these values around may be 
       desirable in order to allow certain quantities to be plotted, but
       where convergence is faster if these values are discarded before 
       the hyperparameters are updated.

   sample-variances

       Samples values for the case-specific noise variances in a 
       regression model where these vary (equivalently, where the noise 
       has a t distribution).  This operation requires that the latent
       values for the training cases be available.  If these are not
       recorded, such values are automatically sampled and then discarded.

Tempering methods and Annealed Importance Sampling are not currently
supported.

            Copyright (c) 1995-2004 by Radford M. Neal