[R] Automatic differentiation (was: Re: D(dnorm...)?)

Alberto Murta amurta at ipimar.pt
Thu Jan 26 14:07:47 CET 2006


On Thursday 26 January 2006 07:56, Bill.Venables at csiro.au wrote:
> While symbolic computation is handy, I actually think a more pressing
> addition to R is some kind of automatic differentiation facility,
> particularly 'reverse mode' AD, which can be spectacular.  There are
> free tools available for it as well, though I don't know how well
> developed they are.  See:
>
> http://www-unix.mcs.anl.gov/autodiff/AD_Tools/
>
> I admit this is not quite the same thing, but for statistical
> computations this is, in my experience, the key thing you need.  (Well,
> for frequentist estimation at any rate...)
>
> There are commercial systems that use this idea already, of course.  Two
> that I know of are 'ADMB' (and its associated 'ADMB-RE' for random
> effects) estimation and of course the 'S-NUOPT' module for another
> system not unlike R.
>
> ADMB is, frankly, difficult to use but it performs so well and so
> quickly once you get it going nothing else seems to come close to it.  I
> has become almost a de-facto standard at the higher end of the fishery
> stock assessment game, for example, where they are always fitting huge,
> highly complex and very non-linear models.
>
> Bill V.
>

I think AD Model Builder is mainly used for fisheries assessment in North 
America and, it seems, also in Australia. In Europe, R is still the de-facto 
standard for fisheries assessment. However, I'd like to support Bill 
Venables' suggestion. I've been resisting to adopt AD model builder, or to 
start using again that other system not unlike R, mainly because of the 
licence price and because I really like R as a tool for almost everything. 
But an AD function would really make a huge difference for my work. There are 
free tools that can be used to perform AD on C or Fortran code (e.g. 
http://www.autodiff.org). One of the difficulties to use them with R is the 
need to translate the R code into C of Fortran code, but probably there are 
many other problems that I'm not able to see.

Alberto
-- 
  Alberto G. Murta <amurta at ipimar.pt>
National Institute of Agriculture and Fisheries Research
Avenida de Brasilia 1449-006 Lisboa Portugal
Tel: +351 213027120 | Fax: +351 213015948




More information about the R-help mailing list