[ADMB Users] The benefits of ADMB

[Arni Magnusson]

Laurie raised a few points besides DLL compiling:

> ... a worked example of an ADMB/R package with a discussion of benefits 
> would be useful eg. The benefits of using ADMB rather than R non-linear 
> solvers (e.g. optim/minpack.lm), the benefits of using R for graphics, 
> data-bases, and linking with other tools. Then a discussion of object 
> orientated programming in R, i.e. S4 classes and FLR.
>
> I think a candidate for this example could be Pella-T example.

The benefits of using ADMB rather than R nonlinear solvers like 
optim/minpack.lm are mainly speed, reliability, and built-in features like 
random effects and MCMC.

Benchmark results like http://otter-rsch.ca/SchnuteRichardsOlsen1998.pdf, 
http://otter-rsch.com/admbre/examples/logistic/logistic.html, and 
http://otter-rsch.com/admbre/examples/sdv/sdv.html indicate that ADMB is 
faster than Gauss, Matlab, S-Plus, and BUGS. I agree with Laurie, that 
given the current variety of nonlinear solvers available in R, it would be 
interesting to benchmark them against ADMB using a Pella-Tomlinson model.

If you want random effects and/or MCMC inside R, the functions are 
limiting compared to ADMB. R packages like nlme, lme4, MCMCglmm, MCMCpack, 
and mcmc are practical for many things, but can only express a small 
subset of the models that can be built with ADMB. I have found this 
especially true for models that are time-structured but not basic 
time-series models.

R is excellent for many things, but there are some tasks that are better 
performed by more specialized software, like databases or spreadsheets, or 
a programming language with particularly strong support for specific 
tasks, like ADMB and Perl. Even expert R programmers recognize its 
limitations, and delegate some tasks to other languages, including 
C/C++/Fortran to crunch numbers, and Perl to build packages.

If I was to do a Pella-Tomlinson assessment, I would probably start with a 
quick Excel prototype, to get some initial parameter values and to verify 
later that I get the same numbers in ADMB. Then I would write the model in 
ADMB (probably borrow from the existing pella-t.tpl example), diagnose and 
plot the model fit in R, run ADMB -mcmc, diagnose and plot the MCMC 
results in R, evaluate possible transformations and reparametrizations, 
and then plot and summarize the final results in R. Even though R is my 
preferred programming language, I would use Excel and ADMB for the things 
they do better. There are many other candidate tools, but these are the 
ones I would use.

The benefits of ADMB become more important when dealing with models with 
hundreds of parameters and very large datasets.

I have only used optim() and other nonlinear solvers in R to satisfy my 
curiosity, never for work. The approach is very similar to ADMB: prepare 
data objects, define parameters, make predictions, and evaluate the 
objective function. Even the code will look quite similar, but in my 
opinion I have a more reliable and versatile tool if I implement a model 
in ADMB.

Of course, it would be nice to back up this opinion with a Pella-Tomlinson 
benchmark. To test the "reliable" part, we would need a difficult dataset 
with local minima. Does anyone know of a suitable dataset?

Arni

P.S. I'm looking forward to trying out the FLR 
(http://r-forge.r-project.org/R/?group_id=318) way to do Pella-Tomlinson 
and other assessment models. The FLash package is particularly interesting 
to ADMB users, as it implements automatic differentiation for other FLR 
packages using ADOL-C, although it appears not quite ready for testing. 
FLR is an ambitious project, and even if the S4 classes take some of the 
playfulness out of R, they're probably the best approach to manage a 
project of this scope.