Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!panix!zip.eecs.umich.edu!umn.edu!stingray!pesc0002
From: pesc0002@stingray.micro.umn.edu (EDWARD S PESCHKO)
Subject: A simple but daunting question
Message-ID: <D2tIrF.Kqz@news.cis.umn.edu>
Sender: news@news.cis.umn.edu (Usenet News Administration)
Nntp-Posting-Host: stingray.itlabs.umn.edu
Organization: University of Minnesota, Minneapolis: Under Grad Workstation Lab.
Date: Sun, 22 Jan 1995 17:48:07 GMT
Lines: 31

hey all --

I needed some practical advice on how to do the following:

I have a series of (somewhat complicated, unknown 2D functions) that I want to 
model... what I was thinking of doing (as a training methodology) is:

Start with a couple of data points, and train the neural net from those few 
data points.
Add more and more datapoints -- and train from there.

I *know* that this is possible because I have seen stock predictions, insurance
predictions, etc. that do exactly this... but I am only familiar with static
problems like character recognition, where the nodes are static:

ie:
(static # of INPUT NODES) --- (static # of HIDDEN NODES) -- (static # of OUTPUT NODES)

wheras I am looking more for:

(dynamic # of INPUT NODES) -- (dynamic # of HIDDEN NODES) -- (dynamic # of OUTPUT NODES)

or some sort of way to 'fit' the dynamic problem into a static network.

(any solution -- simple or complex -- on how to do this is welcome...)

Also, what is the best model of NN to use for this task?

Thanks much,

Ed
