Discussion about R capabilities seem to suggest that for optimization
one may want to use some other software. As I've been working (with Ravi
Varadhan mainly) to try to improve what is now called optim(), I needed
to test some new methods, including a variant of CG. Coding only in R on
a pretty vanilla 3Ghz PC, I was surprised that a generalized Rosenbrock
function in n=50000 parameters solved in less than 2 minutes. It seems a
lot of my experiences in building routines that appear in optim() --
which were written on a machine with 8K (that's K) bytes for program AND
data -- are really not valid any more. This has prompted us to try (and
it not simple) to set up some infrastructure to get good (well, better?)
measures of performance for optimization and nonlinear parameter
estimation. Complications involve the variety of environments and
configurations, as well as codings of test functions, choices of how to
provide gradient information etc. Contact me off-line if you are
interested in this, as we would like it to be relatively easy to use and
share. That can come only by communication, and we want to have lots of
"real" tests, and they take a fair bit of effort to set up in a
standardized way.
However, the main message here is to ask, as I was reasonably asked by
another R worker, that people be much more cautious with conjectures. We
can and should check our opinions by measuring, no matter what tasks or
tools, when we give advice.
JN
