Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!newsfeed.pitt.edu!gatech!newsfeed.internetmci.com!EU.net!sun4nl!cs.vu.nl!newshost.vu.nl!surfnet.nl!spider.research.kpn.com!usenet
From: Joost Hage <hage@cs.utwente.nl>
Subject: Re: calculating the fitness of an ANN
Content-Type: text/plain; charset=us-ascii
Message-ID: <311613D8.5916@cs.utwente.nl>
Sender: usenet@news.research.kpn.com (USENET News System)
Nntp-Posting-Host: gd8117.research.kpn.com
Content-Transfer-Encoding: 7bit
Organization: KPN Research, The Netherlands
Mime-Version: 1.0
Date: Mon, 5 Feb 1996 14:27:36 GMT
X-Mailer: Mozilla 2.0b6 (Win16; I)
Lines: 20

Kamp B. wrote:

> As I understand, a NN is good if performance on both training and test
> set are high. Couldn't you just sort the found performances of the
> trained nets on the average of training and test performance, and then
> eliminate those nets with a high standard deviation between both
> performances (or another measurement of a critical distance between
> both performances (for example minimum (treshold) levels)?But, when using the test-set as fitness-criterium for your GA, you 
implicitly train on the test-set as well. Therefore you might get a 
network that performs very well on both the trainings-set as the test-set 
(and thus has a high fitness), but still generalizes very poorly.

> Bart Kamp
> b.kamp@kub.nl 

-- 
	Hmm...let me think	(Kohonen)
~~~|
   |oost  Hage (hage@cs.utwente.nl)
\_/
