Skip to content

nmle: gnls freezes on difficult case

3 messages · Brian Ripley, Nicholas Lewin-Koh

Original
Nicholas Lewin-Koh
# 
Hi,
Following up on my own post, if in gnlsControl I specify
opt='optim' gnls exits quite nicely, with an error, which
is what I would expect. Is this a bug in nlminb?

Nicholas


tt<-gnls(response~SSllogis(conc,A,B,xmid,scal),tdat,start=start,weights=varPower(),verbose=TRUE,control=gnlsControl(opt='optim'))
**Iteration 1
GLS step: Objective: 29.04254varStruct  parameters:
    power 
0.3373206 

NLS step: RSS =  0 
 model parameters:-0.799941  8.99983  -0.522623  212.314  
 iterations: 2 

Convergence:
   params varStruct 
 1.172208  1.000000 
Error in optim(c(coef(gnlsSt)), function(gnlsPars) -logLik(gnlsSt,
gnlsPars),  : 
  initial value in 'vmmin' is not finite
In addition: Warning messages:
1: In log(xmid) : NaNs produced
2: In log(xmid) : NaNs produced
3: In log(xmid) : NaNs produced
Brian Ripley
#
On Thu, 18 Oct 2007, Nicholas Lewin-Koh wrote:

            

      
      
      
    

        
Possibly, but complex algorithms can loop for non-bug reasons.  I don't 
feel motivated to explore it in detail, but if you want/are able to, 
please let us know what you find.

I have always found optim to work better in package nlme than nlminb (and 
I know others who have the same experience).

            

      
      
      
    

        

      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Nicholas Lewin-Koh
    

    

      
      #
    

  


  

      
Hi,
Thank you for the pointer, I will make the default in my code optim.
As I specified in my last post it may be my installation. In this
case the problem is that the third parameter (xmid) goes negative and 
the log(xmid) fails in the parameterization I am using. For some
reason nlminb is not exiting gracefully. I can look more into this
later if I get time.

Ideally I should constrain xmid to be strictly positive, how would
I use L-BFGS-B in gnls? gnls does not accept a ... argument. In the 
gnlsControl documentation it says

optimMethod: character - the optimization method to be used with the
          'optim' optimizer. The default is '"BFGS"'.  An alternative
          is '"L-BFGS-B"'.

but there is no obvious way to specify lower and upper. Either an
example if there is some scoping trick to do it would be nice,
or change that sentence to "optim method 'L-BFGS-B' will not work
since there is way to specify the upper and lower constraints"

Thanks

Nicholas


Nicholas

On Thu, 18 Oct 2007 15:46:01 +0100 (BST), "Prof Brian Ripley"
<ripley at stats.ox.ac.uk> said: