R nls results different from those of Excel ??

This thread unfortunately pushes a number of buttons:

- Excel computing a model by linearization which fits to
     residual = log(data) - log(model)

rather than
     wanted_residual = data - model

The COBB.RES example in my (freely available but rather dated) book
at http://macnash.telfer.uottawa.ca/nlpe/  shows an example where
comparing the results shows how extreme the differences can be.

- nls not doing well when the fit is near perfect. Package nlmrt is 
happy to compute such models, which have a role in approximation. The 
builders of nls() are rather (too?) insistent that nls() is a 
statistical function rather than simply nonlinear least squares. I can 
agree with their view in its context, but not for a general scientific 
computing package that R has become. It is one of the gotchas of R.

- Rolf's suggestion to inform Microsoft is, I'm sure, made with the sure 
knowledge that M$ will ignore such suggestions. They did, for example, 
fix one financial function temporarily (I don't know which). However, 
one of Excel's maintainers told me he would disavow admitting that 
"Bill" called to tell them to put the bug back in because the president 
of a large American bank called to complain his 1998 profit and loss 
spreadsheet had changed in the "new" version of Excel. Appearances are 
more important than getting things right. At the same conference where 
this "I won't admit I told you" conversation took place, a presentation 
was made estimating that 95% of major investment decisions were made 
based on Excel spreadsheets. The conference took place before the 2008 
crash. One is tempted to make non-statistical inferences.

JN