funfits (PR#137)

ripley@stats.ox.ac.uk ripley@stats.ox.ac.uk
Wed, 10 Mar 1999 07:29:47 +0100


On Wed, 10 Mar 1999 novo@uiuc.edu wrote:

> Hi,
> 
> I spotted an inconsistency between the output of funfits under R, more
> specifically "nnreg", and the one obtained under S-plus. Notice the difference
> in the "Root MSE", "GVC" and "GVC cost=2". As a check mark, I am also including
> the data summaries, confirming that I was in fact using the same data set.
> Here's my output from both systems:

Yes, but does this problem have a unique solution?  Only rarely do neural
net fits not have multiple local optima. More specifically, on the basis of
the figures here it is the S-PLUS output that is wrong as the fits under R
are achieving lower minima. Note (I have not checked here) that often R is
working in double precision where S-PLUS uses floats (and this is Fortran
code, so this is likely).  In my own neural nets package, that did change
the answers using the same compiler and same version of S-PLUS: it is a
function of the ill-posed nature of the problem.  I also get different
answers with the same code and different compilers.

I do think that problems with a package should be taken up with the
author/porter:

auk% less DESCRIPTION 
Package: funfits
Version: 2.0-3
Author: S original by Doug Nychka <nychka@stat.ncsu.edu>
        R port by Albrecht Gebhardt <albrecht.gebhardt@uni-klu.ac.at>.

If you feel the need to file a report to R-bugs on a package, we do need an
example that we can reproduce and work on.  On the evidence so far I am
not convinced there is a bug here.

-- 
Brian D. Ripley,                  ripley@stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272860 (secr)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595


-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-devel mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To: r-devel-request@stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._