[R] A problem about "nlminb"

[Ravi Varadhan]

Popo,

If you indeed have 200000 unknowns to be estimated, I would suggest that you check out spg() function in the "BB" package.  This requires small storage and hence can better handle high-dimensional problems.

Ravi.

____________________________________________________________________

Ravi Varadhan, Ph.D.
Assistant Professor,
Division of Geriatric Medicine and Gerontology
School of Medicine
Johns Hopkins University

Ph. (410) 502-2619
email: rvaradhan at jhmi.edu


----- Original Message -----
From: spencerg <spencer.graves at prodsyse.com>
Date: Saturday, May 30, 2009 4:57 pm
Subject: Re: [R] A problem about "nlminb"
To: David Winsemius <dwinsemius at comcast.net>
Cc: r-help <r-help at r-project.org>


>       You example is NOT self contained, which means that any 
> potential 
>  respondent must guess what you mean by "a function with a variable of 
> 
>  almost 200,000".  The following clarifies this: 
>  
>  
>   > start0 <- rep(1, 200000)
>   > msLE2 <- function(x)sum(x^2)
>   > nlminb(start=start0, msLE2, control = list(x.tol = .001))
>  Error in vector("double", length) : vector size specified is too large
>  
>  
>        "traceback()" reveals that this error message was generated in 
> by 
>  'vector("double", length)', where length = 130 + (n * (n + 27))/2), 
> and 
>  n = length(start) = 200,000 in this case.  This is 20e9 double 
> precision 
>  numbers or 160 GB.  This suggests you need to rethink what you are 
>  trying to do. 
>  
>  
>        In my opinion, in any problem with more than a fairly small 
> number 
>  of unknowns, e.g., 3 or 12 depending on the complexity of the 
> problem, 
>  the vast majority of the unknowns will be better estimated by 
>  considering them as different samples from some abstract population 
> and 
>  trying to estimate first the hyperparameters of that population and 
> then 
>  the individuals conditioned on the hyperparameters.  The most general 
> 
>  tools for that kind of thing in R are in the 'nlme' and 'lme4' 
>  packages.  To understand those, I highly recommend Pinheiro and Bates 
> 
>  (2000) Mixed-Effects Models in  S and S-PLUS (Springer).  If your 
>  observations can not reasonably be considered by mixed-effects models 
> 
>  with normal errors, a second reference is Gelman and Hill (2006) Data 
> 
>  Analysis Using Regression and Multilevel/Hierarchical Models 
> (Cambridge 
>  University Press).  If neither of those seem adequate to your 
> problem, I 
>  suggest you consider using the "RSiteSearch.function" in the 
> RSiteSearch 
>  package to look for other capabilities in R related to your 
> particular 
>  application. 
>  
>  
>        Hope this helps. 
>        Spencer Graves    
>  
>  
>  David Winsemius wrote:
>  >
>  > On May 30, 2009, at 2:19 PM, popo UBC wrote:
>  >
>  >> Hello everyone!
>  >>
>  >> When I use "nlminb" to minimize a function with a variable of 
> almost 
>  >> 200,000
>  >> dimension, I got the following error.
>  >>
>  >>> nlminb(start=start0, msLE2, control = list(x.tol = .001))
>  >> Error in vector("double", length) : vector size specified is too large
>  >> I had the following setting
>  >>
>  >> options(expressions=60000)
>  >> options(object.size=10^15)
>  >
>  > That would do nothing on my machine, but then you may have a 
> different 
>  > (unspecified) OS. You may have unrealistic expectations. 10^15 
> seems a 
>  > bit optimistic to me, even if you were supplying that number in a 
>  > manner that R would recognize.
>  >
>  > ?mem.limits   #  should give you information specific to your OS.
>  >
>  > If you use Windoze, try also:
>  >
>  >  
>  >
>  >
>  >  
>  >
>  >
>  >>
>  >> I have no idea about what might be wrong. Any suggestion is highly
>  >> appreciated!!
>  >
>  > And we have no idea what sort of setup you have. You could, of 
>  > course,  read the specifics for your OS in the Installation Guide:
>  >
>  > cran.r-project.org/doc/manuals/R-admin.pdf
>  >
>  
>  ______________________________________________
>  R-help at r-project.org mailing list
>  
>  PLEASE do read the posting guide 
>  and provide commented, minimal, self-contained, reproducible code.