NOTES ON THE VERSION OF 1997-01-18: Changes in this version: 1) Programs for handling Gaussian process models have been added. 2) Yet more new Markov chain Monte Carlo methods have been added. These may or may not be of any use for neural network models. 3) A "bin" directory is now present, containing symbolic links to all the programs. You can therefore put just this directory in your search path, rather than all the directories. 4) The 'n' option is now allowed for net-pred when the data model is 'class'. It gives the mean probability for each class, and a single squared error figure that is the sum of the squares of the differences between these probabilities and the 0/1 indicators for the class. 5) The 'a' and 'b' quantities are now defined for 'class' models. The give the absolute and squared errors of the vector of class probabilities with respect to the vector of indicators in which there is a 1 for the correct class and 0s elsewhere. 6) The introductory documentation has been reorganized a bit to accomodate the new Gaussian process facilities. A new example on regression with outliers has been added. 7) A file "make.include" has been added to the main directory, which is included in all the makefiles. This allows compiler options, etc. to be set in one place. Bug fixes: A) A bug regarding computation of predictive means when the targets were log transformed has been fixed. B) A bug that might have produced an incorrect error message when specifying an array quantity has been fixed. C) A few other minor bugs have been fixed as well. Portability: a) Log files are now opened as binary files; this should make no difference on Unix systems, but is the proper way to do things. b) The coding for the CPU time feature has been changed to use the right symbolic constant (CLOCKS_PER_SEC), and to try to bypass the overflow problem as well as possible. c) The code no longer relies on M_PI being defined in <math.h>. d) The random number generation module (rand.c) has been changed to read the file of natural random numbers (randfile) one byte at a time, in order to avoid assuming that an 'int' is four bytes long. This will cause the random number stream for a given seed to be different from that of the last release on a machine with little-endian byte order (which shouldn't be a problem to most people). e) As far as I know, log files created with the previous version of 1996-08-26 should still be readable with the new version, on the same machine as they were created on. Known bugs and other problems in this release: 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 "sample-values" and "scan-values" operations for Gaussian processes always recompute the inverse covariance matrix, even when an up-to-date version was computed for the previous Monte Carlo operation.