Professional Basis of AI Backprop Hypertext Documentation
Copyright (c) 1990-99 by Donald R. Tveter
Bugs and Recent Changes
I've never kept track of bugs and recent changes but now on August
10, 1998 I'm going to start:
- October 14, 1999: Freezing the net wasn't working for
rprop and supersab, that's now fixed.
- October 11, 1999: Improved training speed when
there are units turned off.
- August 28, 1999: Polished off the details of the
Unix/Linux/DOS non-GUI version.
- August 28, 1999: Documented my experimental version of
SARPROP that I'm calling XSARPROP at least for the time being.
- June 19, 1999: Updated and expanded recur.html, added
sample problems waves.bp, rshort.bp and rshort2.bp.
- November 6, 1998: Added the Winston animal identifying
example, see the files animals.bp and animals.dat
- October 29, 1998: Added the 12-2-12 encoder problem as an
- October 29, 1998: Changed the code so that there are fewer
error messages when the program gives the "ran out of unit
- October 29, 1998: Changed the t command so it will give
out fewer error messages when there is no network.
- October 29, 1998: A preliminary version of SARPROP was
added, it increases the probability of convergence on some
problems and gives better test set results on others. Its not
documented yet since I'm not completely sure this version is
final. To get the SARPROP algorithm specify a temperature
greater than 0 on the RPROP (R) menu. Reasonable values for
the temperature are 0.01 to 1.
- August 10, 1998: before this version input taken from a file
(i filename) would not catch lines longer than the limit of 255.
- August 8, 1998: before this version input taken from a file
(i filename) would give an unexpected text error if the data
extended across more than one line.
Also around April 3 I found a bug in the recurrent network
code that made it unable to read in recurrent network patterns
that use the h symbol. FIXED.
On March 17 I found a bug where the program would not read in
a program generated .bp file completely. FIXED.
While compiling a new version as C++ using gcc the compiler
turned up a bug where a register variable defaulted to
int rather than being REAL. This can cause mistakes in
classification problems with online updates and the off by
1 option turned on but only on the training set. FIXED.
- The useless but slightly enlightening tr and trp commands
are not working.