Reposted with pdf figure instead of too big scg

JPD



-------- Message original --------
Sujet:  Optimization
Date :  Thu, 15 Nov 2012 18:27:47 -0500
De :    Jean-Pierre Dussault <[email protected]>
Pour :  [email protected]



Hi all,

I am preparing examples for an optimization course for students in image science. I use an example from http://www.ceremade.dauphine.fr/~peyre/numerical-tour/tours/optim_1_gradient_descent/ to promote the use of better algorithms than the simple gradient descent.

I attach the convergence plot of the norm of the gradient for 5 variants of the optim command: gc unconstrained, gc with bounds [-%inf,%inf], gc with bounds [0,1], gc with bounds [0,%inf] and nd. I also include the gradient descent.

Except for the [0,%inf] variant, the solution has all components strictly in [0,1] as displayed here:

   
-->[max(xoptS),max(xoptGC),max(xoptGCB),max(xoptGCBinf),max(xoptGCB0inf),max(xoptND)]
     ans  =

0.9249840 0.9211455 0.9216067 0.9213056 1.0402906 0.9212348

   
-->[min(xoptS),min(xoptGC),min(xoptGCB),min(xoptGCBinf),min(xoptGCB0inf),min(xoptND)]
     ans  =

0.0671743 0.0718204 0.0678885 0.0714951 0.0772300 0.0714255

On the convergence plot, we clearly see that the gradient norm of the gc with [0,1] bounds stalls away from zero while with no bounds or infinite bounds, it converges to zero. This is even more severe for the variant with bounds [0.%inf], which no more approaches the solution, making virtually no progress at all after some 30 function evaluations.

Is it a Scilab bug or a bad example for the gcbd underlying routine? The cost function is strongly convex of dimension 65536. Has someone experienced a similar behavior?


This is unfortunate since I wish to convince my students to use suitably constrained models instead of enforcing constraints afterward.

Thanks for any suggestion to work around this troublesome situation.

JPD












Attachment: Comp.pdf
Description: Adobe PDF document

_______________________________________________
dev mailing list
[email protected]
http://lists.scilab.org/mailman/listinfo/dev

Reply via email to