nnbugs.doc

基于MATLAB完成的神经网络源程序 大家

                           NNUTILS.ZIP ver 1.01
           Gregory Stevens ([e-mail])

                             BUG DOCUMENTATION

If you find any bugs, crashes, or have any fix or feature implementations, 
please send me info on what you have/found at

                     [e-mail]

What I have found so far for the general public to deal with is in the 
following:

NNWHAT:  Okay, this works perfectly when compiled under UNIX.  However, 
apparently the size of the structure is too big for DOS, or maybe it was just
too big for Borland C++.  Either way, it was seizing up if I used anything 
other than the huge memory model, it was then not able to quit (memory
allocation error in command.com), and would, from time to time, over-write
other variables in calls, so that variables would periodically change when,
in the code, no changes were being made to them.  Weird stuff.  I hate
memory errors. 
   I think this may be a DOS limitation, but I don't know.

NNSIM2:  Okay, this algorithm is supposed to discover its own correlations
within the input sets, and divide them up into as many divisions as there are
output nodes.  However, in the experiment which should have correlated the
data into input patterns with greater numbers of nodes in each of the four
quadrants, it only iused three of the four output nodes EVER.  The first
output node is consistently not being used.  I have the feeling this is a
stupid typo error in the code, but I haven't found it yet.



No other bugs were found, but just as a reminder: I plan on releasing more demo
files for unsupervised and recurrent nets in the future.

If anyone wants to implement special combinations, like an unsupervised
recurrent net or something, with these utilities, email me and I will include
to with your credit info in the next release.

greg

[e-mail]