[Coin-ipopt] Re: ipopt question

[Andreas Waechter]

Hi Zhexin

> My application involves hundreds of thousands of variables, thus memory is a
> critical issue. Is it possible to use only single instead of double
> precision in IPOPT?  Besides, do you think ipopt can handle NLP problem of
> that size?

I think that using single precision is not a good idea at all, since it
reduces numerical accuracy significantly, which I expect will be very
hurtful within an interior point algorithm where one encounters rather
ill-posed linear system etc.

We have used Ipopt to solve problem with up to a few million variables,
but it is of course impossible to give a general answer to your question.
It depends very much on how dense the constraint Jacobian and the Hessian
of the Lagrangian function are, how much fill-in is encountered in the
factorization (those factors determine memory and CPU time requirements),
and also on factors such as "degree of nonlinearity" and degeneracy (which
determines whether the code will converge in a reasonable amount of
iterations).  If in general your Jacobian and Hessian matrices are very
sparse, you could try an evaluation of the performance of Ipopt on your
model with AMPL (which is probably easier than hooking it up to your C or
Fortran program), and if it seems to work Ok, you could hook it up to your
code and see if it scaled up.

Regards,

Andreas

PS: Please use the Ipopt mailing list to ask general questions about
Ipopt.

On Fri, 29 Jul 2005, Xiang, Zhexin (NIH/CIT) wrote:

> Dear Andreas,
>
> My application involves hundreds of thousands of variables, thus memory is a
> critical issue. Is it possible to use only single instead of double
> precision in IPOPT?  Besides, do you think ipopt can handle NLP problem of
> that size?
>
> Best,
>
> zhexin
>