Re: non linear regressions and histograms.

[Jean Bréfort <jean brefort normalesup org>]

Le mardi 10 janvier 2006 à 09:10 +1100, msevior physics unimelb edu au a
écrit :
> > Hi,
> >
> > I have made some work on non linear regression curves in graphs, and
> > especially in histograms. As a sample, I implemented a gaussian fit. The
> > patch is available, for interested people, at:
> > http://jean.brefort.free.fr/histograms-regs.patch (~30kB).
> >
> > There are probably other distributions that it might be interesting to
> > implement in the same way (any suggestion?).
> >
> > I do not know if R2 can be evaluated for such a regression, and even
> > what it would mean, so it is not enabled in that case.
> > I also fixed go_non_linear_regression to make it work. BTW, should we
> > keep the same code in gnumeric (current situation) or should gnumeric
> > use the functions in go-regression.c?
> >
> > Waiting for your comments and suggestions.
>
> Hi Jean,
>         This is great! Physicists often need to fit complex functions to
> data.
>
> However these are often so complex that they really need to be compiled
> against a library. I was wondering if it were possible to write a really
> well documented plugin that uses "C-minuit"
> http://c-minuit.sourceforge.net/
> so that sophisticated users could code their own fitting routine.
>
> Just a crazy idea.

Not so crazy, or we are two ;-)

Thanks for the link, I was not aware of it. That said,
go_non_linear_regression seems to have the same capabilities (although
this needs some testing, IMHO).
Writing a plugin like the gaussian fit one is quite easy. It took me
about four hours, because I did not know how go_non_linear_regression
and because this function was buggy (it did not update the parameters),
so I had to fix it. A couple of hours should be enough, including some
test time. What I can easily do is write a tutorial on how to develop
such a plugin and provide templates to make things easy.

A more crazy idea would be to write a plugin taking the source of the
function as parameter... It is probably possible, but, in the short
term, there are so many other things to implement, that I'll have no
time to spend on that.

Cheers,
Jean