Re: DM: Genetic Algorithms

[Warren Sarle]

I think John did not send this to the list, so I am copying his entire
message.

John Aitchison <jaitchison@acm.org> wrote:
>
> Warren wrote , in response to Ron
> 
> > Ron Hartman asks:
> > > There's been a lot of "buzz" about using Genetic Algorithms in 
>Data Mining Applications.
> > > 1.) Can anyone share their experiences using these techniques 
>to solve business problems, especially in direct marketing 
>applications?
> > > 2.) Are they ready for "Prime Time"?  (are there any 
>significant risks?)
> > > 3.) What are good resources to further my knowledge?
> > 
> > GAs are very fragile and require a large amount of tuning and
> > customization by the user; otherwise they will fail to solve even
> > extremely simple problems (Jennison and Sheehan 1995).  I have 
>seen many
> > papers that show that GAs are better than other poor algorithms, 
>or are
> > better than good algorithms with bad initial values. For example, 
>there
> > are numerous papers showing that GAs work better than standard 
>backprop
> > for neural nets, but practically _anything_ works better than 
>standard
> > backprop.  But I have never seen any demonstration that GAs are 
>better
> > than other good algorithms with reasonable initial values for any 
>models
> > commonly used for data mining.
> > 
> >    Jennison, C. and Sheehan, N. (1995), "Theoretical and Empirical
> >    Properties of the Genetic Algorithm as a Numerical Optimizer,"
> >    Journal of Computational and Graphical Statistics, 4, 296-318.
> > 
> 
> I think it is fair to point out that the Jennison and Sheehan paper 
> restricted its attention to performance as a numerical optimizer 
>and 
> makes passing reference to the  possibility of "other uses" of GAs. 

That's true, but many data mining methods do depend on numerical
optimization, including conventional statistical methods such as
logistic regression, as well as trendier methods like neural nets
and empirical decision trees.

> In addition they did find that some modified/hybrid versions of the 
> GA approach can produce satisfactory results. The general 
>conclusion 
> being that there appeared to be some cause for concern about the 
>use of 
> GA's in optimization problems but that the jury was still out.

I agree, and I have heard rumors of situations where GAs perform well,
but I haven't seen the evidence yet.

> A recent paper in "Pattern Analysis and Machine Intelligence" (Nov 
> 96, vol 18, no 11) "Exploring the Power of Genetic Search in 
>Learning 
> Symblic Classifiers" by Neri and Saitta, appears to show that GA's 
>do moderately well in 
> building classification systems for large and complex data sets, 
>but 
> the differences in performance between the various approaches 
> studied were not large.

Thanks for the reference.

> Perhaps the best conclusion is that there is no silver bullet, and 
>it 
> might be of interest to know the sorts of "business problems" that 
> Ron envisaged possibly applying GA's to.
> 
>  
> John Aitchison <jaitchison@acm.org>
> Data Sciences Pty Ltd
> SYDNEY, AUSTRALIA
> 


-- 

Warren S. Sarle       SAS Institute Inc.   The opinions expressed here
saswss@unx.sas.com    SAS Campus Drive     are mine and not 
necessarily
(919) 677-8000        Cary, NC 27513, USA  those of SAS Institute.
* Do not send me unsolicited commercial, political, or religious 
email *