[Coin-ipopt] How to use subroutine EV_HLV(...)

Andreas Waechter andreasw at watson.ibm.com
Mon Aug 1 13:29:28 EDT 2005

Hi Danh,

As I wrote in an earlier message today to Zhexin, it is impossible to make
predictions about this without knowing a lot more details about the model.
The 6GB might be enough, if your matrices are really sparse, and if the
fill-in in the linear system is not very large.

As for avoiding writing code for second derivatives:  Ipopt has
quasi-Newton options, with perform Ok, but there is still the issues that
the larger the number of degrees of freedom is for your problem (ablut
240,000 in your case), the more iterations are usually required (as a rule
of thumb), whereas the iteration count usually grows only slowly with the
dimensionality of a problem when exact Hessian information is used.
Particularly when solving such a large problem, I would recommend to code
the second derivatives, unless they are very complicated or impossible to
compute (e.g. if the functions are obtain via a black box simulations or
so), or when the exact Hessian is not sparse.



On Mon, 1 Aug 2005, Nguyen An Danh wrote:

> Hi Andreas,
> I am preparing to solve the problem of:
> N=960 001 + 1 280 000 (slack variables) = 2 240 001 variables
> M= 1 280 000 + 1 205 000= 2 485 000 constraints
> We are using AIX (IBM), 6.00 Gb. Does default version of Ipopt is enough?
> In my honour, I wish that default version of Ipopt is enough for my
> problem so that I must not make any code for Hessian.  :-)
> Danh.

More information about the Coin-ipopt mailing list