[SciPy-user] Non-negative least squares method

dmitrey dmitrey.kroshko@scipy....
Thu Apr 10 14:42:42 CDT 2008


Gael Varoquaux wrote:
> On Thu, Apr 10, 2008 at 08:24:23PM +0300, dmitrey wrote:
>   
>> Afaik scipy & numpy has no the one, mb bvls from scikits.openopt can be 
>> helpful.
>> Regards, D.
>>     
>
> Can't this be done with a cobyla (constraint optimisation, can be found
> in scipy.optimise) ?
>   
Yes, of course, but cobyla is not specialized solver for the LLS 
problem. scipy lbfgsb could be much more appropriate, since it can 
handle derivatives, that are easy to compute here.

However, even the one fails to solve rather large-scale problems (yields 
sufficiently worse obj func value), I had mentioned the issue here: 
http://openopt.blogspot.com/2008/03/new-llsp-solver-bvls.html
If you have OO installed, you can see /examples/llsp_2.py for benchmark 
scipy_lbfgsb (or any other OO NLP solver, btw I hadn't ALGENCAN checked, 
mb it works better) vs bvls, try different N = 10, 50, etc.

D.


More information about the SciPy-user mailing list