Parameters setting in functions optimization

Wed, Nov 30, 2011 7:16 AM

optimx does allow you to use bounds. The default is using only methods from optim(), but
even though I had a large hand in those methods, and they work quite well, there are other
tools available within optimx that should be more appropriate for your problem.

For example, the current version of optimx should work quite well with lower and upper
bounds specified, and you can see which methods work well by putting
control=list(all.methods=TRUE). all.methods would be overkill for general use of course.

I am hoping to have a new version of optimx up on R-forge within a week -- there are so
many options to check -- that traps the NaNs etc. if it can, and also allows parameter and
function scaling as well as several other new features. This is all experimental at the
moment, but initial results are promising. This is less about "better methods" than about
trapping errors and bad scaling etc. However, if you are able to share your script and
data, I'll be happy to use it as a test and report back to you  if you can communicate it
to me off-list.

Best, JN

On 11/30/2011 06:00 AM, r-help-request at r-project.org wrote:

Message: 68
Date: Tue, 29 Nov 2011 19:15:43 +0100
From: Diane Bailleul <diane.bailleul at u-psud.fr>
To: r-help at r-project.org
Subject: [R] Parameters setting in functions optimization
Message-ID: <4ED5214F.7030504 at u-psud.fr>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed

Good afternoon everybody,
I'm quite new in functions optimization on R and, whereas I've read 
lot's of function descriptions, I'm not sure of the correct settings for 
function like "optimx" and "nlminb".
I'd like to minimize my parameters and the loglikelihood result of the 
function.
My parameters are a mean distance of dispersion and a proportion of 
individuals not assigned, coming from very far away.
The function LikeGi reads external tables and it's working as I want 
(I've got a similar model on Mathematica).

My "final" function is LogLiketot :
LogLiketot<- function(dist,ms)
{
res <- NULL
for(i in 1:nrow(pop5)){
     for(l in 1:nrow(freqvar)){
res <- c(res, pop5[i,l]*log(LikeGi(l,i,dist,ms)))
     }
         }
return(-sum(res))
             }

dist is the mean dispersal distance (0, lots of meters) and ms the 
proportion of individuals (0-1).
Of course, I want them to be as low as possible.

I'd tried to enter the initials parameters as indicated in the tutorials :
optim(c(40,0.5), fn=LogLiketot)

 >Error in 1 - ms : 'ms' is missing

But ms is 0.5 ...

So I've tried this form :
optimx(c(30,50),ms=c(0.4,0.5), fn=LogLiketot)
with different values for the two parameters :
                     par  fvalues      method fns grs itns conv KKT1 
KKT2 xtimes

 >2    19.27583, 25.37964 2249.698        BFGS  12   8 NULL    0 TRUE

TRUE   57.5

 >1 29.6787861, 0.1580298 2248.972 Nelder-Mead  51  NA NULL    0 TRUE

TRUE   66.3

The first line is not possible but as I've not constrained the 
optimization ... but the second line would be a very good result !

Then, searching for another similar cases, I've tried to change my 
function form:

LogLiketot<- function(par)
{
res <- NULL
for(i in 1:nrow(pop5)){
     for(l in 1:nrow(freqvar)){
res <- c(res, pop5[i,l]*log(LikeGi(l,i,par[1],par[2])))
     }
         }
return(-sum(res))
             }

where dist=par[1] and ms=par[2]

And I've got :
optimx(c(40,0.5), fn=LogLiketot)
                     par  fvalues      method fns grs itns conv KKT1 
KKT2 xtimes

 >2 39.9969607, 0.9777634 1064.083        BFGS  29  10 NULL    0 TRUE

NA  92.03

 >1 39.7372199, 0.9778101 1064.083 Nelder-Mead  53  NA NULL    0 TRUE

NA  70.83
And I've got now a warning message :

 >In log(LikeGi(l, i, par[1], par[2])) : NaNs produced

(which are very bad results in that case)


Anyone with previous experiences in optimization of several parameters 
could indicate me the right way to enter the initial parameters in this 
kind of functions ?

Thanks a lot for helping me !

Diane

-- Diane Bailleul Doctorante Universit? Paris-Sud 11 - Facult? des Sciences d'Orsay Unit?
Ecologie, Syst?matique et Evolution D?partement Biodiversit?, Syst?matique et Evolution
UMR 8079 - UPS CNRS AgroParisTech Porte 320, premier ?tage, B?timent 360 91405 ORSAY CEDEX
FRANCE (0033) 01.69.15.56.64