sci.op-research

On Feb 16, 2:06 am, Sebastian Nowozin wrote:
> Hello,
>
> On Feb 16, 12:27 am, Hans Mittelmann wrote:
>
> > On Feb 15, 2:44 pm, hmob...@gmail.com wrote:
> > > Any recommendation for a good *free* solver?
> > So, you insist on Matlab?! That's bad. Let me look again....
> > Are you a student? If yes, contact MOSEK.
>
> I don't think MOSEK qualifies as _free_ solver.
>
> Recently, Peter Carbonetto wrote an excellent Matlab interface to the
> COIN IpOpt interior point solver. It is included in the official
> distribution of COIN IpOpt available athttps://projects.coin-or.org/Ipopt
>
> I used it for convex programming and the performance is excellent (and
> reported results in the literature are on-par with KNITRO). However,
> the memory demand depends a lot on the sparsity structure of the
> problem, ymmv. Also, the robustness and performance is crucially
> dependant on which linear system solver is used, see the
> recommendations on the IpOpt page.
>
> As for the Matlab linprog/quadprog solvers, from my experience I
> recommend against using them at all, even for small scale toy
> problems. They are seriously broken and I wonder how these solvers
> can actually be sold as part of the Optimization toolbox, it seems no
> effort went into principled testing during development as even
> standard benchmark problems cause the solver to exit errorneously.
>
> Sebastian

I agree with Sebastian's recommendation but he is wrong about MOSEK.
They do something for students. Also, I am assuming that the user has
larger than student version size problems. Otherwise one could
generate Matlab interfaces to solvers such as LOQO etc