[R] Looking for constrained optimisation code

[Hong Ooi]

_______________________________________________________________________________________


Ah! I didn't notice this until now.

Yes, following Nocedal & Wright (excellent book, that) I've defined a
quadratic penalty function to enforce an equality constraint, and added
it to the objective. (Originally I had an inequality constraint in mind,
but the equality one doesn't run into problems when you're on the
boundary -- as is likely to happen in this case -- and gets the job
done.) Getting the gradient right was a bit tedious, but it seems to be
working now.

I did find an interesting thing in working with both Splus and R,
though. In Splus, both optim and nlminb run fairly slowly with a pure-S
objective function supplied, even when it was as vectorized as I could
make it. So I thought to compile the objective function in C, and this
sped it up by a factor of 2-4x. On the other hand, the pure-S version
actually ran quite fast in R, and the compiled version actually _slowed_
it down by a small but noticeable amount.

Is this a common experience? Is the R interpreter so efficient that
using compiled code is unlikely to improve speed further? (I'm using R
2.2 and Splus 7.04, on a Windows XP workstation with 3GB of memory.)


-- 
Hong Ooi
Senior Research Analyst, IAG Limited
388 George St, Sydney NSW 2000
+61 (2) 9292 1566
-----Original Message-----
From: Spencer Graves [mailto:spencer.graves at pdf.com] 
Sent: Tuesday, 29 November 2005 12:58 PM
To: Hong Ooi
Cc: Thomas Lumley; r-help at stat.math.ethz.ch
Subject: Re: [R] Looking for constrained optimisation code

	  Have you considered migrating the constraints into the
objective 
function, then cranking up the penalty for constraint violation once you

have a more or less feasible solution?

	  spencer graves

Hong Ooi wrote:

>
________________________________________________________________________
_______________
> 
> 
> You know, this is the first time I've heard of constrOptim.
>  
> I actually have a rather complicated, nonlinear boundary expression in
> mind, so this function by itself isn't quite what I'm after. Still, I
> should be able to hack up a barrier function in my own code and feed
> that into optim/nlminb/constrOptim.
> 
> Thanks!
> 
> 

-- 
Spencer Graves, PhD
Senior Development Engineer
PDF Solutions, Inc.
333 West San Carlos Street Suite 700
San Jose, CA 95110, USA

spencer.graves at pdf.com
www.pdf.com <http://www.pdf.com>
Tel:  408-938-4420
Fax: 408-280-7915



_______________________________________________________________________________________

The information transmitted in this message and its attachme...{{dropped}}