R help - Aug 2011 - Scaling problem in optim()

I am using the function optim and I get the error message
ABNORMAL_TERMINATION_IN_LNSRCH. Reason for this could be a scaling problem.
Thus, I used parscale in order to scale the parameters. But I still have the
error message. For example, with parscale=c(rep(1,n), 0.01,1,0.01):


return(optim(c(mu1,b,k,phi), neg2loglikelihood, method = "L-BFGS-B",
               lower=c(rep(-1,n),1e-5,1e-5,1e5)
,upper=c(rep(5,n),1.29,100,1.9),control=list(parscale=c(rep(1,n), 0.01,1,0.01))
               )
         )
>$par
 [1]  0.9684324  0.9618763  0.9632684  0.9667272  0.9596587  0.9661914
 [7]  0.9731076  0.9684477  0.9723313  1.1679275 92.3677202  0.4947495

$value
[1] -5438.992

$counts
function gradient 
     177      177 

$convergence
[1] 52

$message
[1] "ERROR: ABNORMAL_TERMINATION_IN_LNSRCH"

By using small values for the last parameters, for instance
parscale=c(rep(1,n),0.001,0.9,0.001), there is no Error message anymore, but
then the parameters (n+1) and (n+3) stay at the starting values and are not
optimized anymore.

I have tried several different values and combinations for parscale, but either
I get the error message ABNORMAL_TERMINATION_IN_LNSRCH or at least one parameter
is not optimized. But I want to have both. What can I do? I haven?t specified a
gradient function. I am grateful for any suggestions! Thanks you!


--

Rebekka Schibli <rebekka_schibli <at> gmx.ch> writes:
> 
> I am using the function optim and I get the error message
ABNORMAL_TERMINATION_IN_LNSRCH. Reason for this
> could be a scaling problem. Thus, I used parscale in order to scale the
parameters. But I still have the
> error message. For example, with parscale=c(rep(1,n), 0.01,1,0.01):
> 
> return(optim(c(mu1,b,k,phi), neg2loglikelihood, method =
"L-BFGS-B",
>                lower=c(rep(-1,n),1e-5,1e-5,1e5)
> ,upper=c(rep(5,n),1.29,100,1.9),
> control=list(parscale=c(rep(1,n), 0.01,1,0.01))
>                )
>          )
> 
> >$par
>  [1]  0.9684324  0.9618763  0.9632684  0.9667272  0.9596587  0.9661914
>  [7]  0.9731076  0.9684477  0.9723313  1.1679275 92.3677202  0.4947495
> 
   From these ending parameters it looks as though

parscale=c(rep(1,n),1,100,1)

  would be a better choice?

If you can specify a gradient that will probably help.
Beyond that, we'll probably have to have a reproducible example
in order to provide further help.

  Ben Bolker

There is an error in your call to `optim'.  The`lower' bound is
incorrect:

lower=c(rep(-1,n),1e-5,1e-5,1e5)

This should be:

lower=c(rep(-1,n),1e-5,1e-5,1e-5)

I am not sure if this fixes the problem, but it is worth a try.

I do not understand your scaling.  From the lower and upper bounds it seems like
(n+2)th parameter is the only one that needs to be scaled by a scale factor of
100.

After you have fixed the error in `lower' and the scaling, you might also
want to try "optimx" package that runs a bunch of optimization
algorithms.

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@jhmi.edu<mailto:rvaradhan@jhmi.edu>


	[[alternative HTML version deleted]]