home *** CD-ROM | disk | FTP | other *** search
- This package contains 4 exectuable files of the GMU bp simulator: bps.unix.exe,
- bpS8088.exe, bpS286.exe, and bps.sun360 . The first is for a UNIX system --
- we are running it under ULTRIX 3.0 on a VAX 8530. The second is a plain
- 8088 DOS machine with no math coprocessor, the third
- is for 80286/80386 DOS machines, and the 4th is a Sun 3/60 port. This
- last seems to work OK, but is not presently fully verified. I am in the
- process of updating to Version 1.01, which corrects a bug that is important
- only if you use the "symmetric" version of the sigmoid transfer function --
- the derivative was not calculated correctly for this. At present only the UNIX
- and Sun executables are updated -- network problems & workload have delayed
- uploading the PC versions.
-
- There is a user tutorial, that you may not really need
- to study, called TUTOR.xxx; it comes in 2 flavors also:
- TUTOR.mac is in Microsoft Word 4.0 format for Macintosh
- computers and TUTOR.ASC is an ASCII version, text plus
- line breaks, of the same thing. The latter doesn't
- have the diagrams or formatting of the former.
-
- Files beginning XOR are sample network, training, and
- production files: XOR.231 is a 2-3-1 backprop network
- that has been trained on the XOR.DAT training file, and
- XOR.PRO contains input data to test the trained network.
-
- Source code licenses are now available in return for a fee of $50.00
- US, payable in advance ( we have no way to generate billing). The fee
- gives you right to use and modify the source code at your site, and
- comes with code documentation. The code is about 8K lines of highly
- modularized and "plain vanilla" C code -- porting from MS-DOS to Unix
- required changing 2 lines in the random number generator to obtain
- a system clock value for the seed.
-
-
- To familiarize yourself, just run the simulator and interact
- with the screens. The important parameters are eta -- learning
- rate-- and alpha -- momentum. There is a backprop speedup
- technique called gradient correlation (IJCNN Jan '90) that
- compares the direction of the current and last gradient
- vector, and either proceeds with the backprop step [ifthe
- 2 grads point in essentially the same direction] or discards
- the current weight aadjustment, reduces eta & alpha, and tries
- again. All parameters have default values that are user-resettable.
-
-
-
- Eugene Norris
- CS Dept George Mason University Fairfax, VA 22032 (703)323-2713
- enorris@gmuvax2.gmu.edu
-
-