[R] Powell's unconstrained derivative-free nonlinear least squares routine, VA05AD

[David Kinniburgh]

I have used Mike Powell's optimization routine (VA05AD) from the Harwell Subroutine Library (HSL) for more than 20 years. It is no exaggeration to say that it has helped make my career (thanks Mike). I recently learned that I am not alone in this respect - apparently it still has a loyal following in all sorts of fields!

It is an exceedingly fine piece of software - fast, reliable and easy to set up. On some early problems that I looked at, it was twice as fast as the equivalent NAG routine.  To quote from the manual,  "A hybrid method is used combining features from the Newton -Raphson, Steepest descent and Marquardt methods and calculating and maintaining an approximation to the first derivative matrix using the ideas of Broyden." 

Now that I have converted to R, I will miss my trusted friend. I have started using nls() but have not accumulated enough experience to compare the two. It would be great if VA05AD could be an option in there (algorithm="Powell"). To this end, I recently enquired of the custodians of the HSL whether it would be possible to make it freely available to the R community. The answer was basically 'yes'. Would anybody be willing to port it across? I am not sufficiently capable of doing this yet.

HSL Archive: http://www.cse.clrc.ac.uk/nag/hsl/contents.shtml 

===================
Dr D G Kinniburgh
British Geological Survey
Crowmarsh Gifford
Wallingford
OX10 8BB 
UK

Phone: 0044 1491 692293
Fax:      0044 1491 692345
email:   dgk_at_bgs.ac.uk