Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!newsfeed.pitt.edu!gatech!news-out.communique.net!mr.net!europa.clark.net!cpk-news-hub1.bbnplanet.com!news.bbnplanet.com!feed1.news.erols.com!news.ecn.uoknor.edu!news.ysu.edu!odin.oar.net!malgudi.oar.net!utnetw!news
From: dweddin@uoft02.utoledo.edu
Subject: Backprop NN...what ratio of training to test data?
Message-ID: <E6t01s.7L8@utnetw.utoledo.edu>
Lines: 29
Sender: news@utnetw.utoledo.edu (News Manager)
Organization: University of Toledo
Date: Sun, 9 Mar 1997 15:28:24 GMT



hey all,

I'm using a backprop net for the first time.

I'm using the net for classifying data points into 1 of two sets.

There are 190 vectors in my data set. I plan, of course, to use
some of these for training and some for testing.

What ratio should I use?

A friend of mine who works with BPNN says that I should train
on 70% and test on 30%.

My experience with RBF's says just the opposite. I train RBFs on
smaller sets. 

Anybody got any experience that they want to share?



Is there a good rule of thumb?


Don Wedding


