Hi!
Iñaki Ucar writes:
> It's the same check, but using ATLAS as the algebra back-end instead of
> BLAS.
I wasn’t aware of these differences. I just found the paragraph in the R
Installation and Administration guide that talks about the accuracy of
different implementations causing different solutions.
I think in my case the data used for the test might be the real culprit
because it is a kind of a corner case. Some implementation detail seems
to cause a significant difference in the result (6% off for one
coefficient). Other tests using different data sets do not fail, so I
will look into this.
> It seems that you are checking exact output values in your tests, and
> changing the library changes the computed coefficients. This is generally a
> bad idea, precisely as this issue demonstrates. Checking ranges or numbers
> with some tolerance is much better, but then you cannot rely on .Rout.
I already reduced the number of digits to print because I had this issue
before. I think the computed coefficients changed from one R release to
the next in the past some time ago...
Maybe I should switch to something like
stopif(all.equal(...)
instead.
Thanks again!
Regards,
Stefan
__
R-package-devel@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-package-devel
