It's possibly the L in L-BFGS-B that is more interesting for some
problems, namely for the "Limited Memory", so running without bounds
can
make sense. Unfortunately, the version of L-BFGS-B in R is from the
1990s, and Nocedal et al. released an update in 2011. Maybe someone will
want to work on putting that into the distributed R. I'll be happy to
kibbitz off-list to see what can be done.
The types of algorithms that compete for low-memory (i.e., when you have
1000s of parameters) are the CG codes -- Rcgmin is better than CG by a
long shot for most problems, and I was involved in both -- and truncated
Newton methods, of which I only know of one version in nloptr that in
the very limited tests I've done seems not to be as speedy as I'd
anticipate (there's also an lbfgs in that package -- not tried). I'm
slowly working on getting an all-R truncated Newton code up so different
options could be tried.
If it's the B you need, then Rcgmin, Rvmmin (essentially optim:BFGS)
work with gradients, as does nlminb, while dfoptim:nmkb or dfoptim:hjkb
or minqa:bobyqa (also bobyqa in nloptr, again I've no more than tried
one example) are available for no-derivative case, and possibly some
other routines.
JN
On 13-08-02 06:00 AM, r-help-request at r-project.org
wrote:> Message: 36
> Date: Fri, 02 Aug 2013 10:38:51 +0200
> From: Uwe Ligges<ligges at statistik.tu-dortmund.de>
> To: Anera Salucci<a.salucci at yahoo.com>
> Cc:"r-help at r-project.org" <r-help at r-project.org>
> Subject: Re: [R] optim with(out) box constraints!
> Message-ID:<51FB701B.70201 at statistik.tu-dortmund.de>
> Content-Type: text/plain; charset=ISO-8859-1; format=flowed
>
>
>
> On 02.08.2013 10:10, Anera Salucci wrote:
>> >Hello R users,
>> >
>> >Does optimizing a function using optim with method=
"L-BFGS-B" and without box constraints lead to "L-BFGS"
optimization in R?
> Sort of, but the question is why this would be beneficial with today's
> computers ...
>
> Best,
> Uwe Ligges
>
>
>