From: wsilvert@netcabo.pt
I came late to this discussion, having been away for a while, but I find the
topic surprising. I did some work with a colleague from the Netherlands on
using neural networks to generate fuzzy memberships, using as a training set
the assignments of a team of experts, so I find the two approaches, fuzzy
logic and neural networks, complementary rather than competitive. Thus some
of the material posted seems perfectly reasonable, some doesn't!
About Dmitry's comment that neural nets are not tunable, it surprises me
that no one has addressed this limitation. I recall once attending a talk
where a NN was used to identify people from photographs, and with a training
set of 10,000 photos it could make correct identification whether or not the
subjects wore glasses. I asked whether the NN could be told to ignore
glasses, which would reduce the training set to 5,000 pictures, but was told
that couldn't be done. Why not?
As for whether expert systems require training sets, I would ask where the
experts get their expertise. It seems to me that the expertise can be viewed
as the result of processing the data in what could be considered a training
set. If we could teach a NN to ignore eyeglasses it would become a bit more
like an expert.
Bill Silvert
"Dmitry A. Kazakov" wrote in message
news:bul2ovo2a611qjm4i1e1ph7tuehsb6d7uv@4ax.com...
> >> Neural networks provide a way to represent and organize data. Fuzzy
> >> logic is an extension of the conventional logic. What is to compare
> >> here? Absolutely nothing! A valid comparison could be neural vs. fuzzy
> >> neural networks. But it would be just crisp vs. fuzzy.
> >>
> >Actually, neural networks can provide a large number of functions,
> >including (for example) classification. Thus for certain problems one
> >can use either a neural net or a fuzzy expert system. They are, of
> >course, quite different techniques.
>
> I would disagree. A fuzzy expert system could be built on the basis of
> a neuronal network. I mean, as the knowledge carrier one could take a
> network instead of a data base of rules. From this point of view a
> valid comparison could be: rules data base vs. neuronal network. And
> again the word "fuzzy" have slipped away! (:-))
>
> >An advantage of neural nets is that little or no a priori knowledge is
> >required;
>
> This only means that the learning algorithm is not tunable. It has no
> parameters. Whether it is an advantage, is another question.
>
> > the corresponding disadvantage is that after the neural net
> >is constructed and tuned, one has little or no idea how it reaches its
> >conclusion. There has been a lot of work into extracting a rule base
> >that corresponds to the neural net from the neural net connections,
> >but I have the impression that the results of this work are not
> >terribly satisfactory.
>
> Me too.
>
> >However, I am not up to date on this work, so
> >what I say may not be correct.
> >
> >A disadvantage of neural nets is that a training set of data and the
> >corresponding conclusions is abolutely required. Some applications
> >cannot possible meet this requirement.
>
> After all it is a learning with a teacher, so it has its
> disadvantages, like a necessity to have that "teacher".
>
> >Expert systems do not usually require a training set of data for their
> >construction; instead, expert kowledge is used to construct the rules.
>
> Yes, but it again, it is comparing apples and oranges in my view.
>
> >They do, of course, require calibration of model parameters and
> >subsequent validation on real-life data. Neural nets also require
> >validation.
> >
> >This is a very short treatment of the subject, but is perhaps better
> >than nothing. I know of no publication that deals with the topic.
>
> ---
> Regards,
> Dmitry Kazakov
> www.dmitry-kazakov.de
--- SoupGate-Win32 v1.05
* Origin: you cannot sedate... all the things you hate (1:229/2)
|
