[SciPy-user] a small OpenOpt example howto benchmark solvers

dmitrey dmitrey.kroshko@scipy....
Wed Nov 21 13:57:04 CST 2007

Nils Wagner wrote:
> On Wed, 21 Nov 2007 21:11:32 +0200
>   dmitrey <dmitrey.kroshko@scipy.org> wrote:
>> Hi Nils,
>> I suppose I have provided all required URLs.
>> Let me mention the ones once again:
>> blog report about results
>> http://openopt.blogspot.com/2007/11/example-of-oo-nlp-solvers-benchmark.html
>> and this one contains URLs to code (7th line from top, 
>> "the example"):
>> http://projects.scipy.org/scipy/scikits/browser/trunk/openopt/scikits/openopt/examples/nlp_bench_1.py
> Sorry for the confusion. I was talking about a reference 
> from the literature.
> Nils
That example is not from literature, it's just almost 1st that I decided 
to code and bench (looks very similar to nlp_2.py from 
/openopt/examples/, but objfunc is (x-M)**1.5, not **2, to prevent 100% 
pure quadratic func).
Almost all examples from literature available for me are too small to be 
taken into openopt time and cputime comparison, because (as I have 
mentioned you in a letter) openopt requires some time for preparation 
prob struct before and after calculations; also, stop criteria mean too 
much for so small problems.
As you see from the example, ALGENCAN takes much more f,c,h evals (for 
*this* example) time is almost same. It's because the cause I have 
mentioned (it's espesially valid for small problems), + ALGENCAN is 
fortran-written (as well as cobyla), while lincher is 100% Python + 
cvxopt QP solver, that is also mostly Python-written.
Regards, D.

More information about the SciPy-user mailing list