[Coin-ipopt] bounds on a variable

[Andreas Waechter]

Hi Olivier,

I'm not really surprised that the problem is more difficult to solve a 
problem with bounds than without (since you add more constraints).

In the Fortran version, you can try to play with the initial value of the 
barrier parameter, and how the dual variables are initialized.  Also, if 
the starting point is very close to the bounds, the first steps can be cut 
short quite a bit.  To change that, you can either manually change the 
starting point you give to Ipopt, or you can change the values of the 
parameters DBNDFRAC and DBNDPUSH.

In the C++ version, you could try the same things, and in addition, you 
could try to use the "adaptive" choice of the barrier parameter (see the 
parameter mu_strategy).  This tries to choose values for the barrier 
parameter, that allow Ipopt to take larger steps.

Regards,

Andreas

On Mon, 8 May 2006, Olivier Riviere wrote:

> Hello,
> I'm using the fortran version of IPOPT for a problem where some variables
> may at the solution be strictly equal to their lower bounds (indicated in 
> IPOPT). I noticed that the convergence is considerably slowed in regards to 
> the "unbounded" case since it takes time for some variables to tend towards 
> their lower bounds. Does someone knows how to deal with such types of problem 
> ? (I tried to change the value of the parameter dmovebounds but without 
> results...)
> I also notice that the size of the first step performed by the algorithm to 
> reach also seems to depend on the bounds indicated for the variables (if the 
> bounds are large the first step will be large) . Is there a way to control 
> the norm of the first step ?
> Thanks for your help
> Olivier
> _______________________________________________
> Coin-ipopt mailing list
> Coin-ipopt at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/coin-ipopt
>