Re: Advantage

Dmitry A. Kazakov <[EMAIL PROTECTED]> wrote in message news:<[EMAIL PROTECTED]>...

> I would not set off NN against expert systems. Because it is
> theoretically possible to build an expert system on the basis of a NN.
> Consider "experts", which knowledge can be extracted in the form a
> neuronal subnetwork. Then this knowledge, a description of a network,
> is incorporated into a larger system using some standard framework.
> That would be a NN-based expert system.

This is a new idea to me, and is very interesting. The idea of looking
at a neural net that has been conventionally constructed (from a
training data set) as a fuzzy expert system, and extracting the rules
from the resulting neural net, is of course not new; we seem to agree
that this work has not been very satisfactory. But the idea of looking
at a fuzzy expert system that has been constructed from expert
knowledge rather than a training data set is new to me, and strikes me
as being fairly important.

Of course, real-world fuzzy expert systems are a lot more complicated
than simple one-step fuzzy control expert systems. The expert systems
I have constructed ten to be multi-step affairs, with the rules
fireable in one step being fired in parallel, and the results of one
step being fed as input to the next step. I think you mean that we
could look at each epert system step as a layer in a neural net.

We frequently have recursion, with the outputs being fed back as
inputs to steps that have already been fired. Has anyone worked with
neural nets in which the outputs of one layer are fed back as inputs
to the same or a preceding layer?

Sincerely, William Siler