[SciPy-User] least squares producing different answers?

Pauli Virtanen pav@iki...
Thu Jun 14 18:13:37 CDT 2012


14.06.2012 23:40, Laura Matrajt kirjoitti:
[clip]
> Does anyone have any idea if it is indeed the case that leastsq uses
> some type of random initialization?

No, it does not use random numbers.

However, be aware that floating point math with modern CPUs and
compilers is often not reproducible, even across different runs on the
same machine, using the same binary:

http://blog.nag.com/2011/02/wandering-precision.html

http://www.nccs.nasa.gov/images/FloatingPoint_consistency.pdf

I assume the semi-random results in themselves are not a problem for
your case? You can always request a stricter tolerance from the solver
if needed.

-- 
Pauli Virtanen



More information about the SciPy-User mailing list