Otherware/Otherware_1_SB_Development.iso/amiga/programm/language/neural_n.rea

Text File | 1992-11-11 | 2KB | 43 lines

Readme_V3.1st:

     This package contains all the functions necessary to generate a
Neural network which you can train and use in your programs.  This
package contains the C/C++ code for a single hidden layer, and a
2 hidden layer neural network.  The network is feedforward and fully
connected.  You may specify an size for each layer at run-time.
I could have made one Neural_network class that could have the number
of hidden layers specified at run time but it would make the class
interface ugly and somewhat hard to use.  Since most neural networks
(not all) use either one or two hidden layers, this should be enough
for most cases.
     Also included is C++ code to train and test a Time-Delay neural net.
It has two hidden layers and you should see the header file if you
do not know what the structure of a TDNN is.  I do not have time right now
to convert it to C.

This code is Public Domain and anyone can use it in any type of program.
(I hate copyrighted or restricted use class libraries)

     This is version 3 of the C++ Neural network code.  I have made
significant changes to the code.  The names of the classes and a few
functions have been changed.  I added the hybrid delta-bar-delta alg.
and the random optimization algorithm.  Now included is a Time-Delay
neural net class.

    I have tested the Neural_network1 and Neural_network2 classes
extensively and believe there are no major or minor bugs.  I have
somewhat tested the C code and I am pretty sure it is as reliable as
the C++ code from which it was converted.  The Time-Delay NN class I
have not had a chance to seriously test so there may be some bugs in
it but I hope not.

I would like to here from anyone using this code on its performance, any
improvements, options that should be added, or bugs. I will continue to
support both the C and C++ code and make future enhancements as new
algorithms come out.


Charles Anstey
[e-mail]