R help - Aug 2006 - scaling constant in optim("L-BFGS-B")

Hi all,

 

I am trying to find estimates for 7 parameters of a model which should fit
real data. I have a function for the negative log likelihood (NLL) of the
data. With optim(method="L-BFGS-B",lower=0) I am now minimizing the
NLL to
find the best fitting parameters. 

 

My problem is that the algorithm does not converge for certain data sets. I
have read that one should scale the fn (i.e. the NLL in my case), however I
am having trouble with the scaling constant: If I change it, the algorithm
converges for certain data sets, for which it didn't before, but for others
it doesn't converge although it did before. In addition, the scaling
constant affects the value of the optimal parameters and the converging
value of the NLL (evidently). So to be able to compare the parameters
between different data sets I need to use the same scaling constant. Trying
out all values between 0.1 and 1 is very laborious and is not quite a
systematic approach.

 

My question is: Are there any rules of thumb to choose a scaling constant?
And how do I justify it's application (it looks a bit like a "magic
constant" that tricks the algo to converge, but does not have a systematic
justification)?

 

I'd greatly appreciate any hints, tricks or references about the scaling
constant

 

Thanks for your help!

 

Simon

 

********************************************************************

Simon Ruegg

Dr.med.vet.,  PhD student

Institute for Parasitology

Winterthurstr. 266a

8057 Zurich

Switzerland

 

phone: +41 44 635 85 93

fax: +41 44 635 89 07

e-mail: s.ruegg@access.unizh.ch

 


	[[alternative HTML version deleted]]

maybe it would be better to scale the parameters (using 'parscale') 
instead of NLL.


I hope it helps.

Best,
Dimitris

----
Dimitris Rizopoulos
Ph.D. Student
Biostatistical Centre
School of Public Health
Catholic University of Leuven

Address: Kapucijnenvoer 35, Leuven, Belgium
Tel: +32/(0)16/336899
Fax: +32/(0)16/337015
Web: http://med.kuleuven.be/biostat/
     http://www.student.kuleuven.be/~m0390867/dimitris.htm


----- Original Message ----- 
From: "Simon Ruegg" <s.ruegg at access.unizh.ch>
To: <r-help at stat.math.ethz.ch>
Sent: Wednesday, August 09, 2006 2:52 PM
Subject: [R] scaling constant in optim("L-BFGS-B")

> Hi all,
>
>
>
> I am trying to find estimates for 7 parameters of a model which 
> should fit
> real data. I have a function for the negative log likelihood (NLL) 
> of the
> data. With optim(method="L-BFGS-B",lower=0) I am now minimizing
the
> NLL to
> find the best fitting parameters.
>
>
>
> My problem is that the algorithm does not converge for certain data 
> sets. I
> have read that one should scale the fn (i.e. the NLL in my case), 
> however I
> am having trouble with the scaling constant: If I change it, the 
> algorithm
> converges for certain data sets, for which it didn't before, but for 
> others
> it doesn't converge although it did before. In addition, the scaling
> constant affects the value of the optimal parameters and the 
> converging
> value of the NLL (evidently). So to be able to compare the 
> parameters
> between different data sets I need to use the same scaling constant. 
> Trying
> out all values between 0.1 and 1 is very laborious and is not quite 
> a
> systematic approach.
>
>
>
> My question is: Are there any rules of thumb to choose a scaling 
> constant?
> And how do I justify it's application (it looks a bit like a
"magic
> constant" that tricks the algo to converge, but does not have a 
> systematic
> justification)?
>
>
>
> I'd greatly appreciate any hints, tricks or references about the 
> scaling
> constant
>
>
>
> Thanks for your help!
>
>
>
> Simon
>
>
>
> ********************************************************************
>
> Simon Ruegg
>
> Dr.med.vet.,  PhD student
>
> Institute for Parasitology
>
> Winterthurstr. 266a
>
> 8057 Zurich
>
> Switzerland
>
>
>
> phone: +41 44 635 85 93
>
> fax: +41 44 635 89 07
>
> e-mail: s.ruegg at access.unizh.ch
>
>
>
>
> [[alternative HTML version deleted]]
>
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide 
> http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
> 

Disclaimer: http://www.kuleuven.be/cwis/email_disclaimer.htm