[Coin-ipopt] Re: IPopt for Matlab

[Peter Carbonetto]

Salut Julie,

    I'm sending my reply to the IPOPT mailing list as well, since other 
users may have some suggestions. You can subscribe at 
http://list.coin-or.org/mailman/listinfo/coin-ipopt. Also, I recommend 
searching the mailing list archives for further hints.

> I'm using your matlab interface for IPopt and I'm experiencing a few
> problems. I was wondering if you could help.
>
> I guess I have a rather difficult optimization problem with constraints.
> I'm not able to have IPopt converge. When I let it run for too many
> iterations, it gets to an unfeasible point from which it is not able to
> recover. IPopt just exits with an error without giving me the best answer
> found so far.
>
> A similar thing happens when I vary the number of iterations. Once I let
> it run for 1000 iterations and then for 500 iterations. I got a better
> optimum value with the 500 iterations run because, unfortunately, the 1000
> iterations run ended on a worse solution (although a better one had been
> found previously).

This sounds like more of an IPOPT problem than a MATLAB interface problem.
The fact that IPOPT ends up at an infeasible point is very troubling.

Several questions come to mind:

* Do you start at a feasible point?
* Have you tried running the IPOPT derivative checker? Perhaps it is a bug
   in the computation of the gradient or Jacobian.
* Are you computing the Hessian, or using the L-BFGS approximation to the
   Hessian? I've found problems with this approximation for some
   constrained problems.
* 500 iterations seems like a lot. How big an optimization problem is
   this?

Any details you can provide concerning your optimization problem will help 
in pinpointing the problem.

> I'm just wondering if there would be a way to keep track of the best
> optimum value found so far and return in case IPopt doesn't converge.

This shouldn't be necessary, since this should not happen in the first 
place.

Peter Carbonetto
Ph.D. Candidate
Dept. of Computer Science
University of British Columbia
http://www.cs.ubc.ca/~pcarbo