Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!das-news.harvard.edu!news2.near.net!MathWorks.Com!yeshua.marcam.com!usc!sdd.hp.com!heifetz.msen.com!zib-berlin.de!news.th-darmstadt.de!fauern!rrze.uni-erlangen.de!hub-n.franken.de!ark.franken.de!ralf
From: ralf@ark.franken.de (Ralf W. Stephan)
Subject: Re: ALN question: some result
Message-ID: <1994Oct10.141558.1292@ark.franken.de>
Organization: his desk writing an article
Date: Mon, 10 Oct 1994 14:15:58 GMT
References: <1994Oct8.184114.809@ark.franken.de> <3799qe$3p1@scapa.cs.ualberta.ca>
X-Newsreader: TIN [version 1.2 PL2]
Lines: 36

Bill Armstrong writes:
> There is one more thing you could try, which would be extremely simple
> to do.  That is to use the voting option of Atree 2.0.  Boolean nets
> by themselves have generalizing power, but the generalization obtained
> by taking a majority vote of several nets is much greater.  Since the
> training time of the ALN is 0.2 times as long as for BP, perhaps it
> would be a fair comparison to take a majority vote of 5 ALN networks
> to compute each bit of the output. 

You were right.  Here are the numbers:

Misclassifications (of 104), ten test runs (params as before):
---------------------------------------------------------------
3 trees: 17 23 18 27 16 17 15 17 22 20 = 19.2 avg.
5 trees: 16 22 23 13 17 16 14 15 16 15 = 16.7
9 trees: 17 20 16 17 13 10 13 15 12 15 = 14.8
19 trees: 21 13 9 14 12 13 12 13 16 18 = 14.1

Note 1:  There is only one bit of output for the two classes.
         Maybe the results change if we take one bit for every class.

Note 2:  I first got an average of 10 misclassifications, even with
         3 trees!  It turned out an interesting coding error that
         showed that 'a good majority vote limit' need not be 50-50,
         but can also be 40-60.  So there's another parameter to
         play with...
ralf
--
## CrepesSuzettes v9.99 ##
You are in a twisty little article, all maze.



--
## CrepesSuzettes v9.99 ##
You are in a twisty little article, all maze.
