[Ipopt] performance on netlib lp test set

[Andreas Waechter]

Hi Jonathan,

Yes, Ipopt is not specifically written for solving LPs, and in particular, 
it does not do a number of heuristics that interior point solvers for LP 
do (such as finding a good starting point), and the default options are 
geared towards NLPs.  However, as I also wrote in the other posting just 
now, some time ago I had added a "meta option" called "mehrotra_algorithm" 
which, if set to "yes", will activate some options that make Ipopt work 
often better for LPs and QPs.  You might also get some small additional 
speedup by selection the options that tell Ipopt that the derivative 
matrices are constraint (see that other message that I just sent).

Regards,

Andreas

On Fri, 29 Oct 2010, Jonathan Hogg wrote:

> Hi,
>
> I've been running ipopt on the netlib test set using the amplsolver
> interface.
>
> This works, and gets correct answers, however the iteration count is on
> the high side (eg 25fv47 200 iterations). Is this the same experience
> other people have?
>
> I'm told that most netlib problems should be solved by an interior point
> method in more like 20 iterations than 200.
>
> Is this just because IPOPT is designed for NLP, or have I got something
> going wrong?
>
> The .nl file for 25fv47 is available at the following url if anybody
> would like to give it a go:
> http://www.maths.ed.ac.uk/~s0455378/25fv47.nl
>
> Thanks,
> Jonathan.
>
> _______________________________________________
> Ipopt mailing list
> Ipopt at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/ipopt
>
>