[Scilab-users] optim vs Neldermead: improvement
Paul Bignier
paul.bignier at scilab-enterprises.com
Mon Jan 23 11:37:56 CET 2017
Hi Paul,
As I'm sure you've read them, optim
<https://help.scilab.org/docs/6.0.0/en_US/optim.html>, neldermead
<https://help.scilab.org/docs/6.0.0/en_US/neldermead.html> &
numderivative
<https://help.scilab.org/docs/6.0.0/en_US/numderivative.html> help pages
provide you with all the flags you can customize.
After that, it's all in the functions that you provide to compute the
cost & its derivative.
Best regards,
Paul
On 01/23/2017 09:41 AM, paul.carrico at free.fr wrote:
>
> Hi All
>
> I’m using ‘optim’ and ‘NelderMead’ in conjunction with my finite
> element solver.
>
> A “good” optimization is a balance between accuracy and cpu time … in
> other word I do not necessary need to be accurate at 1e-11 but
> requiring a lot of iterations where 1e-3 is enough with a lower amount
> of loops.
>
> In my understanding:
>
> 1. With ‘optim’, I can modifiy
>
> * The step value h in numderivative (put at 1e-3 after previous
> tests on analytical functions tests)
> * The values of epsf (default value) and epsg (tested at 1e-3 and 1e-5)
>
> 1. With “Neldermead”, I can change:
>
> * Tolfunrelative (tested at default value for the moment)
> * Tolxrelative (tested at default value for the moment)
>
>
> Am I right or is there another 'flags' I can play with?
>
> /NB/: so far, Nelder-Mead requires less iterations than ‘optim’ with a
> single variable … I’m wondering how can I improve optim use that is
> supposed to converge faster?
>
> Thanks
>
> Paul
>
>
>
> _______________________________________________
> users mailing list
> users at lists.scilab.org
> http://lists.scilab.org/mailman/listinfo/users
--
Paul BIGNIER
Development engineer
-----------------------------------------------------------
Scilab Enterprises
143bis rue Yves Le Coz - 78000 Versailles, France
Phone: +33.1.80.77.04.68
http://www.scilab-enterprises.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.scilab.org/pipermail/users/attachments/20170123/a3fa7a4b/attachment.htm>
More information about the users
mailing list