[R] L-BFGS-B and parscale in optim()

[Sebastian P. Luque]

On Wed, 25 May 2011 23:46:02 +0000,
Ben Bolker <bbolker at gmail.com> wrote:

> Sebastian P. Luque <spluque <at> gmail.com> writes:
>> Hi,

>> When using method L-BFGS-B along with a parscale argument, should the
>> lower and upper bounds provided be on the scaled or unscaled values?
>> Thanks.

>   I think you can figure this out by experimentation on your own,

I'm asking because I've already experimented and I'm not certain.


> but I believe that all of the scaling done by parscale is internal
> (i.e., you should provide bounds on the original, unmodified scale).

Within the bounds I provide in the original scale I shouldn't find
problems, but optim() reports that my function is returning non-finite
values.  I added a debugging message to the function given to optim() to
see whether those bounds are being violated:

fn <- function(p, l1, l2) {
    res <- -sum(FUNCTION-WITH-p-l1-l2)
    if (any(!is.finite(res))) {
        print(paste("problems with", p, l1, l2))
        } else res
}

which shows me that the estimate for at least one parameter is outside
the bounds.  I'll try to prepare a simplified example to reproduce.


-- 
Seb