[Bug-gsl] different compilers, different results for nonlinear LS fit

[mcdermott]

Dear all,

I am using GSL release 1.4 with both the Microsoft Visual C++ 6.0 (MSVC) 
compiler and the cygwin gcc version 3.2.  Recently I got strange results with 
code I had compiled using MSVC.  I double checked the results against the same 
code compiled using gcc and I got different results.  Moreover, the gcc results 
were correct (at least they agreed with SAS output) and the MSVC seems to have 
the problem.  

The problem seems to be that the two programs go in different directions after 
beginning with the same starting values for the estimates.  This only happens 
about 1 to 2 percent of the time and I can see nothing distinctive about the 
"bad" datasets.  I am fitting a four-parameter logistic curve to sets of 30 
points (3 replicates each at 10 different x values).  Has anyone seen this 
phenomenon before?  I can send code and data if needed.  Thanks in advance.

Regards,
Jim McDermott