Dear Reader concerned in Fuzzy Systems,
I am trying to organise a Session in the Second International Conference
on Knowledge-Based Intelligent Electronic Systems: KES '98, 21st - 23rd April
1998 , Adelaide , Australia (For more details on this conference, please
consult the KES'98 home page: http://www.kes.unisa.edu.au/kes98/kes98.html).
The title of the Session proposed is
"DEVELOPING WELL-PERFORMING FUZZY SYSTEMS FROM NOISY NUMERICAL DATA"
Background:
It is well-known that - like neural networks - many types of Fuzzy Systems
are universal approximators. This is, at least in principle, a promising
result. However, the corresponding theorems do not provide for concrete
methods to find such an approximating system.
On the other hand, many methods have been proposed in order to end up with
a well-performing fuzzy system, among which methods that apply the knowledge
of human experts and methods where fuzzy rules are learned from numerical
data. In the last case, we can choose between a variety of methods like
'direct' methods (e.g., the Wang-Mendel algorithm), methods based on 'genetic
algorithms' and methods based on 'fuzzy neural networks'.
Goal of the Session:
The intention of this Session is to deal with the last group of methods.
It focusses on the question on how to develop well-performing fuzzy systems,
especially in cases where the training data are polluted by NOISE. Central
issues (often inspired by the theory and practice of neural networks) are:
. How well is the approximation by the final system?
. What are the potential causes of a poor approximation, e.g., a wrong
'architecture' of the fuzzy system, a sort of 'local minimum problem of
learning', a wrong 'error function' during training, or something else?
. Are there different systems that perform equally well and can the
simplest one be determined easily (facilitating the fuzzy rule extrac-
tion phase at the end)?
. How should we handle small training sets?
. How do Fuzzy Systems deal with noise?
. Does the problem of overfitting exist?
. If so, which procedures exist to find the model with the 'optimal model
complexity', i.e., the model having the best 'generalization performance'?
Answers to these questions can be of theoretical nature (e.g., by using
mathematical approaches like numerical analysis, statistics, and information
theory) or of practical style (e.g., by using the technique of 'structural
stabilization', 'cross validation' techniques, a 'regularization' approach,
or other heuristics).
Participation:
If you are interested in contributing to this Session, please send me
an e-mail message to the address indicated below. Contrary to the official
KES'98 schedule, I 'll use the following deadlines:
Receipt of papers : 20th October 1997 (at my regular mail address)
Notification of acceptance: 20th November 1997
Receipt of final papers : 20th December 1997
Papers should be written in English (5 to 10 pages maximum). The typeset-
ting information of papers can be found at the afore-mentioned KES'98 home page.
With regards,
Jan van den Berg
__________________________________________________________________________
| regular (snail) mail address:
E-mail: jvandenberg@few.eur.nl | Erasmus University Rotterdam
Phone : +31-10-4081343 | Faculty of Economics
Fax : +31-10-4526177 | Dept. of Computer Science, room H4-23
| P.O. Box 1738, 3000 DR Rotterdam
| The Netherlands
Home Page: http://www.cs.few.eur.nl/few/people/jvandenberg/
__________________________________________________________________________