[SciPy-user] openopt vs. cvxopt, 'f' vs. 'd','z'
Wed Jan 30 13:40:06 CST 2008
Mclean Edwards wrote:
> The numbers turn out to be the same, as I had previously made an error on
> my end. (I needed r.xf[-1] for my optimal value instead of r.ff, due
> to a problem transformation.)
> The speed for openopt-cvxopt and cvxopt are
> also comparable (~15% overhead for a couple of runs on small problems).
As for speed, there is the following issue: cvxopt LP & QP solvers
require a parameter True/False to treat problem sparse or dense. I
decided not to overwhelm OO users additional parameters, moreover,
lpSolve and glpk has no the one, they determine it by themselves,
according to sparsity of matrices. So I decided to call "sparse" CVXOPT
solvers if numberNonZeros/FullSize<0.3, and "dense" otherwise. So the
time elapsed and results can be a little bit different (because other
algs were used).
> Dmitrey, you deserve some praise. Openopt is a good package, and I am
> very glad you are developing it.
Thank you, however, it would be much more better, would anyone mention
something like that in my guestbook - it could help me to achieve a
finance support via grant.
> When I'm finally able to get some
> preliminary code done myself, I would be glad to contribute.
> I'm off to play around some more with the NLP solvers.
More information about the SciPy-user