R help - Feb 2013 - 'gmm' package: How to pass controls to a numerical solver used in the gmm() function?

Hello --

The question I have is about the gmm() function from the 'gmm' package
(v. 1.4-5).

The manual accompanying the package says that the gmm() function is
programmed to use either of four numerical solvers -- optim, optimize,
constrOptim, or nlminb -- for the minimization of the GMM objective
function.

I wonder whether there is a way to pass controls to a solver used
while calling the gmm() function?

In particular, the problem that I have been having is that the gmm()
fails to converge withing the default number of iteration for the
'optim' solver that it uses. Ideally, I would wish to figure out a way
to be able to choose controls, including the number of iterations, for
the solver that I tell gmm() to use.

Currently, the way I call the function is as follows:

model.name <- gmm(g=g.fn, x=data, gradv=g.gr, t0=c(start),
type=c("twostep"), optfct=c("optim") )

I also would want the gmm() function to know that I want it to pass
the following control -- maxit=1500 -- to the optim solver.
Unfortunately, the 'gmm' manual does not tell whether this is doable.

Thanks for your help.

On Feb 19, 2013, at 5:25 PM, Malikov, Emir wrote:
> Hello --
> 
> The question I have is about the gmm() function from the 'gmm'
package
> (v. 1.4-5).
> 
> The manual accompanying the package says that the gmm() function is
> programmed to use either of four numerical solvers -- optim, optimize,
> constrOptim, or nlminb -- for the minimization of the GMM objective
> function.
> 
> I wonder whether there is a way to pass controls to a solver used
> while calling the gmm() function?
> 
> In particular, the problem that I have been having is that the gmm()
> fails to converge withing the default number of iteration for the
> 'optim' solver that it uses. Ideally, I would wish to figure out a
way
> to be able to choose controls, including the number of iterations, for
> the solver that I tell gmm() to use.
> 
> Currently, the way I call the function is as follows:
> 
> model.name <- gmm(g=g.fn, x=data, gradv=g.gr, t0=c(start),
> type=c("twostep"), optfct=c("optim") )
> 
> I also would want the gmm() function to know that I want it to pass
> the following control -- maxit=1500 -- to the optim solver.
The argument name in the manual is `itermax`. I cannot tell from lookng at the
code whether that would get passed to 'optim'.
> Unfortunately, the 'gmm' manual does not tell whether this is
doable.
 There is also a "..." argument which is stated in the help page to be
passed to "optim". Looking at ?optim one sees that controls generally
need to be in a list named 'control'. That this is the intent is
supported by the sentence on page 11 of the gmm vignette:

"We could try dierent starting values, increase the number of iterations in
the control option of
optim or use nlminb which allows to put restrictions on the parameter
space."

-- 

David Winsemius
Alameda, CA, USA

This is a quite general issue that those of us who try to prepare 
optimization tools must deal with quite often. The minqa package 
internal methods were designed to be used with customized controls to 
the algorithm, but we had to "package" them with some more or less OK 
compromise settings. If you use that package directly, you can change 
the values, but not usually if it is wrapped inside something like gmm.

I'd welcome -- off list -- discussions with developers of packages like 
gmm to find a consistent approach to allow the controls to be set by 
those users needing to do so. A danger is that each wrapper package uses 
either its own approach or none at all. My suggestion is that such 
packages include a control() list as an argument with sensible defaults, 
and that this list contains an item (probably also a list) of controls 
to the optimizer(s). A bit ugly, but for most users, not used.

Best,

JN


On 13-02-20 06:00 AM, r-help-request at r-project.org
wrote:
> Message: 2
> Date: Tue, 19 Feb 2013 19:01:35 -0800
> From: David Winsemius<dwinsemius at comcast.net>
> To: "Malikov, Emir"<emalikov at binghamton.edu>
> Cc:r-help at r-project.org
> Subject: Re: [R] 'gmm' package: How to pass controls to a numerical
> 	solver	used in the gmm() function?
> Message-ID:<B54404CA-ACC2-441B-8A1E-B0483695BE7E at comcast.net>
> Content-Type: text/plain; charset=us-ascii
>
>
> On Feb 19, 2013, at 5:25 PM, Malikov, Emir wrote:
>
>> >Hello --
>> >
>> >The question I have is about the gmm() function from the
'gmm' package
>> >(v. 1.4-5).
>> >
>> >The manual accompanying the package says that the gmm() function is
>> >programmed to use either of four numerical solvers -- optim,
optimize,
>> >constrOptim, or nlminb -- for the minimization of the GMM objective
>> >function.
>> >
>> >I wonder whether there is a way to pass controls to a solver used
>> >while calling the gmm() function?
>> >
>> >In particular, the problem that I have been having is that the
gmm()
>> >fails to converge withing the default number of iteration for the
>> >'optim' solver that it uses. Ideally, I would wish to
figure out a way
>> >to be able to choose controls, including the number of iterations,
for
>> >the solver that I tell gmm() to use.
>> >
>> >Currently, the way I call the function is as follows:
>> >
>> >model.name <- gmm(g=g.fn, x=data, gradv=g.gr, t0=c(start),
>> >type=c("twostep"), optfct=c("optim") )
>> >
>> >I also would want the gmm() function to know that I want it to pass
>> >the following control -- maxit=1500 -- to the optim solver.
> The argument name in the manual is `itermax`. I cannot tell from lookng at
the code whether that would get passed to 'optim'.
>
>> >Unfortunately, the 'gmm' manual does not tell whether this
is doable.
>   There is also a "..." argument which is stated in the help page
to be passed to "optim". Looking at ?optim one sees that controls
generally need to be in a list named 'control'. That this is the intent
is supported by the sentence on page 11 of the gmm vignette:
>
> "We could try dierent starting values, increase the number of
iterations in the control option of
> optim or use nlminb which allows to put restrictions on the parameter
space."
>
> -- David Winsemius Alameda, CA, USA