home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.neural-nets
- Path: sparky!uunet!mcsun!Germany.EU.net!infko!inews
- From: evol@infko.uni-koblenz.de
- Subject: Backpercolation vs. other training algorithms
- Message-ID: <1992Nov17.152007.15543@infko.uucp>
- Sender: inews@infko.uucp (inews)
- Organization: University of Koblenz, Germany
- Date: Tue, 17 Nov 1992 15:20:07 GMT
- Lines: 41
-
-
- I am working on a technical report comparing many different training
- algorithms for feedford networks. All algorithms are applied to a
- very hard to solve classification task. Up to now I have applied
- the following algorithms:
-
- 1. Standard Backprop
- 2. Standard Backprop in batched mode
- 3. Standard Backprop in batched mode, learning rate
- set by Harry A. C. Eaton and Tracy L. Olivier's rule
- 4. Standard Backprop with decreasing learning rate
- (Christian Darken and John Moody)
- 5. J. Schmidhuber's adaptation of the step size
- 6. R. Salomon's evolutionary adaptation of the learning rate
- 7. L.-W.Chan and F. Fallside 's adaptation of the learning rate
- by angle observing
- 8. A.H. Kramer and A. Sangiovanni-Vincentelli's Polak-Ribiere
- method with line search
- 9. J. Leoard and M.A. Kramer's conjugate gradient method
- with line search
- 10. F.M. Silva and L.B. Almeida's learning rate adaptation
- by detecting sign changes in the derivatives
- 11. T. Tollenaere's SuperSAB
- 12. M. Riedmiller's RPROP
- 13. Robert A. Jacobs Delta-Bar-Delta technique
- 14. Scott E. Fahlman's Quickprop
-
- I am still looking for some more training algorithms. I have tried to
- implement BACKPERCOLATION. Nervertheless the only report on backpercolation
- I know is not sufficient to implement this algorithm. Does anyone know a
- good paper on backpercolation? Has anyone applied it to bigger problems ?
- Is there a public available implementation (other than SNNS) ?
-
- By the way RPROP and Quickprop are clearly the winners. I will publish
- my report soon (available by ftp).
-
- Thanks in advance
-
- Randolf Werner
-
-
-