Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!news.mathworks.com!uunet!news.nyc.pipeline.com!psinntp!cronkite!blw
From: blw@utrc.utc.com (Brad Whitehall)
Subject: Network Calibration/Scaling:  Help
Message-ID: <1995Apr18.152255.22660@cronkite.res.utc.com>
Sender: news@cronkite.res.utc.com
Nntp-Posting-Host: vega.res.utc.com
Organization: United Technologies Research Center
X-Newsreader: TIN [version 1.1 PL8]
Date: Tue, 18 Apr 1995 15:22:55 GMT
Lines: 34


Hello,

I am looking for work in the area of neural network calibration.  
Calibration might not be the best term, so let me explain what I mean.

I have access to a large amount of data that is "almost" correct.
That is, I can train a network with it.  At a later point, I will
receive a few data points, that are "reality" -- the network should
provide the desired answer for these new points.  What typically happens
is the original data is off from the real model by some factor.  If
it was always just a scaling issue, then I could "shift"
the ANN results to get the desired output, but the desired change
is usually not so simple.

Question: Is there a way to calibrate a network -- learn the general
shape of a function, and then transform it to match a few data points.
My networks have many inputs (multi-dimensional space).

I am interested in references to papers or people working in this area.

I believe it is closely related to the problem of having an on-line learning 
system work in a dynamic environment.  For example, I train my network,
it works great, I put it in the field and it learns over time to adjust 
itself.  Except in my situation, I don't have a lot of time for it 
to learn the proper adjustments.


Please email me any pointers and I will post a summary.  

Thanks for your help,
Brad Whitehall
United Technologies Research Center
blw@antares.res.utc.com
