NOTES ON THE VERSION OF 2004-11-10

Except for the lightly-documented addition of Hamiltonian importance
sampling, this release contains only some fairly minor (though useful)
extensions and some bug fixes.


Changes in this version:

1) Hamiltonian importance sampling was added.  See xxx-his.doc.  More
   tutorial documentation (and maybe a paper!) may be written some day.

2) A "count" data model has been added, which models Poisson count 
   data.  It is implemented only for Gaussian process models, where
   the Gaussian process produces the log of the Poisson mean.

3) The "numin" module has been changed to allow for missing values,
   written as "?", which are allowed for targets by some models (currently 
   only neural network models).

4) The gp-eval program has been extended to allow evaluations at points
   read from a data file, rather than from a grid.  See gp-eval.doc for
   details.

5) An "E" option has been added to net-pred and gp-pred, which gives the 
   expected fraction of wrong guesses at binary targets.  For details,
   see net-pred.doc and gp-pred.doc.  This was documented in the last 
   release, but not actually included in the version of the program 
   that was distributed.

6) Data file reading was changed to not read the whole data file if only
   a part of it is required.  The speed of reading when there are many
   inputs or targets has also been improved.

7) A new "K0" quantity has been added, equal to the expected value of 
   the kinetic energy (the "K" quantity).  See mc-quantities.doc.


Bug fixes.

1) Some out-of-date versions of the copyright notice were replaced with the
   current version.

2) Fixed a bug in dist-est, in which using too many iterations produced
   garbage results, rather than an error message.

3) Fixed a bug that affected "offsets" for neural network models.

4) Fixed some other bugs involving error checking.


Known bugs and other deficiencies.

1) The facility for plotting quantities using "plot" operations in xxx-mc
   doesn't always work for the first run of xxx-mc (before any
   iterations exist in the log file).  A work-around is to do a run of
   xxx-mc to produce just one iteration before attempting a run of
   xxx-mc that does any "plot" operations.

2) The CPU time features (eg, the "k" quantity) will not work correctly
   if a single iteration takes more than about 71 minutes.

3) The latent value update operations for Gaussian processes may recompute 
   the inverse covariance matrix even when an up-to-date version was 
   computed for the previous Monte Carlo operation.

4) Covariance matrices are stored in full, even though they are symmetric,
   which sometimes costs a factor of two in memory usage.

5) Giving net-pred several log files that have different network architectures
   doesn't work, but an error message is not always produced (the results may
   just be nonsense).