GP-GEN:  Generate GP hyperparameters / latent values randomly, or fix them.

Gp-gen writes a series of independently-generated values for the
hyperparameters of a Gaussian process model to a log file.  The
hyperparameters are drawn randomly from their prior, or set to fixed
values.  Optionally, latent values and/or individual noise variances
for each training case are generated as well, from their priors.

Usage:

    gp-gen [ -l ] [ -n ] log-file [ max-index ] [ "fix" [ sc-val rel-val ] ]

Records of hyperparameters with indexes from zero up to the indicated
index are generated (the default is max-index of zero).  If the log
file already contains records with some of these indexes, only records
with indexes greater than the last existing record are generated.

If just max-index is specified, the hyperparameters are generated
randomly from the prior, using the random number seed taken from the
log file (eg, as specified by rand-seed).

The "fix" option is useful for initializing iterative programs.  If it
is given, the hyperparameters are not generated at random, but are
instead set to fixed values.  These values are the top-level widths
for each hyperparameter if no values follow "fix".  If values do
follow "fix", they specify the scale and relevance hyperparameters of
exponential parts of the covariance.  The hyperparameters for the
linear part are set to the product of these two values.  The
hyperparameter for the constant part is set to the scale value given.

No data model is required to use gp-gen, but if a model is specified,
the hyperparameters associated with it are generated as well.

If the -l option is specified, latent values for each training case
are also generated, randomly from the Gaussian process prior defined
by the generated hyperparameters.  If no jitter term is included in
the covariance function, a small amount (1e-6) will be added to the
diagonal of the covariance matrix when generating latent values to try
to avoid numerical problems.

If the -n option is specified, noise variances are generated for each
training case, from their prior given the generated hyperparameter
values.  If the targets are not real valued, or the model does not
have case-by-case noise variances, the -n option is silently ignored.

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