[R] nls convergence problem

[Joerg van den Hoff]

Earl F. Glynn wrote:
> "Berton Gunter" <gunter.berton at gene.com> wrote in message 
> news:007c01c6c0ae$fd260e50$711f210a at gne.windows.gene.com...
>>>   Or, maybe there's something I don't understand about the
>>> algorithm being used.
>> Indeed! So before making such comments, why don't you try to learn about 
>> it?
>> Doug Bates is a pretty smart guy,  and I think you do him a disservice 
>> when
>> you assume that he somehow overlooked something that he explicitly warned
>> you about. I am fairly confident that if he could have made the problem go
>> away, he would have. So I think your vent was a bit inconsiderate and
>> perhaps even intemperate. The R Core folks have produced a minor miracle
>> IMO, and we should all be careful before assuming that they have 
>> overlooked
>> easily fixable problems. They're certainly not infallible -- but they're a
>> lot less fallible than most of the rest of us when it comes to R.
> 
> I meant no disrespect to Doug Bates or any of the R Core folks. I thought 
> what I wrote had a "neutral" tone and was respectful.  I am sorry if anyone 
> was offended by my comments and suggestions.  I am certainly thankful for 
> all the hard work that has gone into developing R.
> 
> efg
> 

well, just a feedback to that: of course the tone of your mail was by no 
means inadequate (at least douglas bates did not object...). the 
tendency on this list to rather harshly rebuke people for some kind of 
(real or imagined) misconduct and to 'defend' R against 'attacks' is 
counterproductive and unnecessary. it goes without saying that the 
people 'behind' R can not and will not (and are not expected to) change 
the code after each mail on the help list which raises some question.

and concerning your `nls' question: sure, the noise requirement is a 
pitfall in the beginning, but afterwards it's irrelevant: you don't fit 
noise free data in real life (in the sense that real data never follow 
you model exactly). and, sure, the convergence decision could be altered 
(given enough time and knowledge). whether convergence failure on exact 
data is a bug or a feature is a matter of taste, probably.

getting better access to the trace output and especially access to 
intermediate pre-convergence values of the model parameters (this would 
  'solve' your problem, too) would really be an improvement, in my mind 
(I think this is recognized by d. bates, but simply way down his 'to do' 
list :-().


joerg