[ADMB Users] an alternative to R for nonlinear stat models

[Arni Magnusson]

I don't use the optimization routines in R. Of course I use canned models 
like lm, glm, gam, lme, etc., but not the free-form optimizers.

Depending on the model, I will often start in Excel to visually fit the 
model, explore the parameter space, get initial parameter values, and a 
feel for the dynamics. After that I take it to ADMB. Coding a model in 
more than one language can also prevent mistakes.

I find Excel useful to sketch something, explore, and explain to others. 
Solver is weak, but seems surprisingly robust. I haven't touched Visual 
Basic for almost a decade, though. If I feel the need for it, it's time to 
move the model to another platform.

I have tried the R optimizers, including recent R packages, to satisfy my 
curiosity. The clunky interface is just a reminder of how elegant the rest 
of R is. What I end up writing is very similar to ADMB, so I might as well 
write it there. Then I get phases, MCMC, speed, and reliability.

If the R optimizers are as unreliable as your results indicate, I'm not 
going to rub it in by cc'ing this to the r-help mailing list. Played 
enough vuvuzela for one day...

Arni



On Wed, 16 Jun 2010, Mike Prager wrote:

> Arni,
>
> I am wondering which optimization routine(s) in R you use. I have had 
> trouble getting even relatively simple models to converge reliably. 
> Sometimes I'll run the optimizer 100 times & pick the best answer. This 
> seems much worse than in the days when I used SAS.
>
> Mike
>