Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!news.mathworks.com!hookup!swrinde!pipex!uknet!cix.compulink.co.uk!usenet
From: jbullard@cix.compulink.co.uk ("John Bullard")
Subject: RE: Question:  Training data with unknown values.
Message-ID: <D3wKyI.C7G@cix.compulink.co.uk>
Organization: Compulink Information eXchange
References: <3fgvqj$a5c@canopus.cc.umanitoba.ca>
Date: Sun, 12 Feb 1995 20:01:29 GMT
X-News-Software: Ameol
Lines: 9

I must confess that I am not an expert, but I've used a back propagation 
network, (Neuroshell) which requests minimum and maximum values for 
inputs.  If your unknown values lie between the two (and they must 
mustn't they?), then you do not have to discard other, known variables. 
All it means is that the more unknown variables in the model, the less 
accurate the model becomes (they model uses the min and max values of 
that variable to forecast/classify outcomes). However, it will not 
degrade the model, it just won't improve it as much as it would have done 
if all the values were present.
