Re: [S] Splus inconsistences in nnet V&R2 example

Prof Brian Ripley (ripley@stats.ox.ac.uk)
Tue, 3 Mar 1998 17:11:58 GMT


Matthew R. Nelson <mrn@superh.hg.med.umich.edu> wrote
>
> A collegue of mine is attempting to apply neural networks in S-PLUS to a
> problem analogous to the example given in V&R2, pp 340-341. He installed
> the latest MASS library, but was unable to reproduce the example results
> when testing the library. Has anyone had similar problems?
>

The only surprise is that he got the same answer on two platforms. I
don't even get the same answer for different compilers (Sun SC3.0,
SC4.0, gcc 2.8.0) on the same platform.

This is not an implementation problem. The neural network fit is very
sensitive to initial conditions. So, for example, is ppreg. All the
answers are sensible ones, but there are many.

The README in the test directory (on Unix) says

Do not be concerned by small differences in the results for sammon.o;
this algorithm is prone to rounding error (it is designed for speed
over accuracy). Similarly, the results for nnet.o and pps.o may
differ even between releases of compilers on the same machine.

I would currently recommend fitting a network twenty times and
averaging the predictions. There is an example of this in chapter 17.

BDR
-----------------------------------------------------------------------
This message was distributed by s-news@wubios.wustl.edu. To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message: unsubscribe s-news