[R-sig-ME] Suggestions for numerical optimization tools...

[Ben Bolker]

  Is optimize() really the bottleneck, or is it the computation time
in the objective function?  Can you implement critical bits of the
objective function calculation in C or FORTRAN?  Based on the
description in optimize(), it seems that although it's a compromise
between robustness and efficiency, that it's usually pretty efficient.
How many function evaluations are typically being required to get
to the minimum?  Can you get away with a lower tolerance?  Can you find
some kind of Gaussian-quadrature approximation to your integral that
makes things more efficient?

  You could also try

http://code.google.com/p/rsympy/#Semi-Automatic_Differentiation

suggested by John Nash ...



H c wrote:
> My current program relies on the ML estimation of a parameter, \phi,  based
> on numerical methods.  The parameter, \phi,  lies on the 0 to 1 interval and
> evaluation of the ML given any value of \phi is computationally expensive.
>  I am currently using the "optimize()" function, optimizing the Likelihood
> with respect to phi.  This is extremely computationally expensive and is the
> bottleneck of an otherwise efficient program.
> Does anyone have any suggestions of better tools for numerical optimization.
>  (Derivatives are not known, so gradient decent options do not appear to be
> applicable).
> 
> Anything helps,
> 
> Harlan
> 
> 	[[alternative HTML version deleted]]
> 
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models


-- 
Ben Bolker
Associate professor, Biology Dep't, Univ. of Florida
bolker at ufl.edu / www.zoology.ufl.edu/bolker
GPG key: www.zoology.ufl.edu/bolker/benbolker-publickey.asc