Re: ITP: iminuit -- Robust Python minimisation library based around MINUIT2

[Drew Parsons <dparsons@debian.org>]

On 2020-08-08 17:10, Jeremy Sanders wrote:
> On 08/08/2020 10:59, Drew Parsons wrote:
> > One question: those benchmarks indicate that scipy CG or Powell 
> > generally performs better.  What is the use-case for minuit? Why not 
> > just use scipy?
>
> In my experience, minimizers do not always run well on all types of
> problems. I've had quite a few problems which get stuck in local
> minima using CG with messy likelihood functions. It's quite useful to
> have different minimizers in one's toolbox, and perhaps to run more
> than one after the other. It's useful to have a well-tested general
> minimizer to also run.


That's fair enough.  In that sense minuit complements scipy/CG, 
providing alternatives for the cases where CG performs below its usual 
standard.


> iminuit also has some useful extra functionality such as computing
> confidence intervals and making contour plots.

These are both very nice features indeed. Quite rare to see confidence 
intervals handled outside the specialised statistics modules.


Thanks Jeremy, good luck with the packaging!

Drew