Hi all,
Here is a summary of substantive replies to my posts (on both s-news and
r-help) regarding use of Excel's Solver for a nonlinear regression problem.
A number of people replied that Solver performs well as an optimizer
based on experience, particularly if given reasonable starting values. A
number of other people replied that it performs poorly based on experience.
Whether or not it arrives at a bona fide solution, several people
pointed out limitations to using Solver associated with lack of
diagnostics and related statistical output. When using Solver, you
cannot "trace the internals" of the solving process, and you cannot
obtain the Hessian matrix or variance-covariance matrix.
One user stated that he sometimes uses Excel's Solver to analyze the
same problem he has previously done in S+ or R; if it arrives at the
same answer, he's happy.
I was also reminded of several general advantages of using a command
line language (ease of debugging, portability to other problems or other
users, etc.). Adding my own comment, that general issue applies to using
Solver in that it is impossible (I think) to tell after the fact what
options were selected for the optimization.
The consensus was that someone in my shoes would do well to learn S+/R,
or AD Model Builder ... but then you're all proficient in S+/R already :-)
Thanks for your input,
Kristian
--
Kristian Omland
Postdoctoral Research Associate
Vermont Cooperative Fish & Wildlife Research Unit
Rubenstein School of Environment & Natural Resources
University of Vermont
Burlington VT 05405
voice: (802)656-2496 fax: (802)656-8683
web page: http://www.uvm.edu/~komland
|