[Coin-ipopt] Best ipopt options for fast QCP solution?

[Andreas Waechter]

Hi Frank,

Yes, the NEOS-AMPL interface provides all required first and second
derivatives to Ipopt.

As for speeding up the execution speed, it's hard to give advice without
seeing how Ipopt behaves on the model.  If the number of iterations are
somewhat high, it might be possible to improve performance by playing with
parameters like dmu0, dfscale, iscale, imuinit (all of which are available
through AMPL).  Also, choosing a good starting point (whatever that
means...) can make a big difference).

If you want you can send me a (preferably small) instance of your problem
where you think the performances could be improved.  (I won't be able to
spend much time with it and I won't promise something comes out of this,
but maybe I see something).  If so, please send the files directly to my
email address (so that we don't spam the mailing list).

By the way, quadratic constraints can be arbitrarily difficult... :)

Cheers!

PS: Thanks for your nice words about NEOS the other day!



On Fri, 4 Mar 2005, Frank J. Iannarilli wrote:

> Hi,
>
>
> I'm using IPOPT, via NEOS-AMPL interface, to solve a
> quadratically-constrained program (QCP).  Since the non-linearity of my
> problem is relatively simple compared to the heavier lifting that IPOPT is
> designed to handle, I surmise there may be options to increase the
> execution speed. Presently, I accept all the default settings.
>
>
> Here's my problem statement:
>
> ==============================================
> Decision variables: x[i], deviation[i]; continuous, >=0, <=bigX
>                   : w[j]; continuous, >=0, <=1
>
> min: deviation[i];  % linear objective
>
> s.t.
>    % Quadratic LAD constraints:
>
>    -deviation[i] <= Sum{j} (a[i,j]*x[i%j]-b[i])*w[j];
>
>    Sum{j} (a[i,j]*x[i%j]-b[i])*w[j] <= deviation[i];
>
>    %%% the notation x[i%j] is shorthand for indexing
>    %%% as some (unstated) function of i and j
>
>    % GUB on w[j]
>    sum{j} w[j] = 1;
> ===============================================
>
>
> Does the NEOS-AMPL interface provide all the derivative/Hessian info to
> IPOPT?
>
> I am willing to run IPOPT locally, if necessary, to obtain the wider array
> of options than available thru the AMPL interface.
>
> Thanks in advance for your suggestions!
>
>
>
> Frank J. Iannarilli, franki at aerodyne.com
> Aerodyne Research, Inc., 45 Manning Rd., Billerica, MA 01821 USA
> www.aerodyne.com/cosr/cosr.html
> _______________________________________________
> Coin-ipopt mailing list
> Coin-ipopt at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/coin-ipopt
>