home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!mcsun!uknet!gdt!brispoly!y_jin
- From: y_jin@csd.brispoly.ac.uk (Y Jin)
- Newsgroups: comp.ai.neural-nets
- Subject: Question: Real Time Local Learning and Global Learning
- Message-ID: <1992Aug18.114005.20649@csd.brispoly.ac.uk>
- Date: 18 Aug 92 11:40:05 GMT
- Organization: Bristol Polytechnic, England
- Lines: 18
-
- Question: Real Time Local Learning and Global Learning
-
- We are applying BP neural networks in nonlinear control problems. The work is
- concentrated on real time learning. Relative references and suggestions will
- be very appreciated. The main idea of our research is presented below:
-
- Supposed that a BP neural network nearly approaches the desired nonlinear
- function through off-line learning. The real time (on-line) learning is used
- to improve the approximation accuracy. The neural network could increase the
- real-time approximating speed locally by increasing the learning rate. However
- by doing so the neural network global approximation features (gained in off-
- line learning) will be damaged. We suggest that the neural network weights be
- split into two parts: linear part for local approximation and nonlinear part
- for global approximation. It is easy to develop a learning algorithm for linear
- part weights, which results in stable control systems. The nonlinear learning
- slowly shifts linear part weights to nonlinear part weights.
-
- Please email ( y_jin@csd.brispoly.ac.uk), I will summarize.
-