Hi all,
I would like to inform you that the fmincon toolbox has been updated
(https://atoms.scilab.org/toolboxes/fmincon/1.0.2) and now includes a
major feature: the Efficient Computation of Sparse Hessian using
Coloring algorithm (scicolpack is now a dependency of fmincon). This
feature was missing and now allows fmincon to consider really big
problems. A example of such is given in the "minimal surface demo" which
compares 3 different approaches of the problem: one with finite
differences gradient, and BFGS Hessian, one with exact gradient and BFGS
Hessian, and the last with exact gradient and finite differences sparse
Hessian. The time factor between the first and the third approchach is
around 200 for a 21x21 square but you can edit the demo file and
increase this value. For a 51x51 square the third approach takes only
0.45 seconds and is 2500 times faster. For a 101x101 square the first
approach takes forever and the third ones takes less than 2 seconds and
is still 10 times faster than the second one.
If some of you still use other softwares for large scale optimization,
maybe you can now give Scilab a try. If you need a better control on
parameters and/or if you are not a former/actual user of Matlab's
fmincon, I recommend using directly ipopt (from sci_ipopt toolbox). But
the fmincon api is quite easy to use. Don't hesitate to took at the code
of the several new demos.
Don't hesitate to contact me if you have some problems using fmincon of
just to say that you are using it with success !
Regards,
--
Stéphane Mottelet
Ingénieur de recherche
EA 4297 Transformations Intégrées de la Matière Renouvelable
Département Génie des Procédés Industriels
Sorbonne Universités - Université de Technologie de Compiègne
CS 60319, 60203 Compiègne cedex
Tel : +33(0)344234688
http://www.utc.fr/~mottelet
_______________________________________________
users mailing list
users@lists.scilab.org
http://lists.scilab.org/mailman/listinfo/users