[Scilab-Dev] optim gc with bounds often fails
Jean-Pierre Dussault
Jean-Pierre.Dussault at USherbrooke.ca
Fri Mar 6 19:32:14 CET 2015
Hi,
I use regularly the optimization functions in scilab. Actually, as a Faculty professor in optimization, I often test prototype optimization algorithms which I implement in scilab and test them. For unconstrained or bound constrained optimization, I compare with the optim command and with L-BFGS-B which I interfaced to be used from scilab.
I also use Scilab in my courses. Unfortunately, the optim command with gc for bound constraints is useless. Therefore, I am forced to instruct the students to install and use L-BFGS-B. One important usage is the denoising/deblurring of images, high dimension bound constrained problems unsuitable for the qn method.
Here is a report in which I give some more details.
https://dl.dropboxusercontent.com/u/18380848/BenchMark.pdf
The concluding table is
qn
unconstrained
for small dimensions
gc
unconstrained
recommended
nd
unconstrained
for non differentiable problems
qn
bounded
for small dimensions
gc
bounded
avoid
In summary, the optim cg for bound constrained problems is useless as it fails to converge most of the time. Perhaps converting professionally the toolbox draft L-BFGS-B I produced would be a useful addition to scilab?
Thanks,
JPD
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.scilab.org/pipermail/dev/attachments/20150306/60ef863d/attachment.htm>
More information about the dev
mailing list