[Coin-ipopt] (no subject)

[Andreas Waechter]

Hi Danh,

> Thank you very much for your very long email.  Your advice is very helpful for
> me.  My problem has run after  I changed a new starting point.  I hope that I
> could have its solution soon.
>
> Could you explain me the following questions?
>
> 1.  Since with the problem of 112001 variables, ma27_call:  LA increased from
> 66993497 to   279769007,  I doubt that another problem of mine which has over
> millions variables will meet difficulty with MA27. (I wish that it would not
> happen!!!) .  So I start to learn how to change MA27 to PARDISO .
> If I did so, what would be advantages and disavantages?  I know that changing
> linear solver has much works.
> What is the disadvantage of PARDISO (Fortran version) on http://
> www.computational.unibas.ch/cs/scicomp/software/pardiso/ ?

It would be quit difficult for you to integrate another solver into the
Fortran version.  Instead, you should probably wait until we have
integrated Pardiso properly into the NEW C++ version of Ipopt (and in the
meantime you might want to switch to the C++ version anyway, since the old
version will not be updated except for crutial bug fixes).

> 2.  Since the mechanical properties of my model, I intend  to add a mechanical
> law  for variables when Ipopt solves equation (9) in your paper "On the
> Implementation of an Interior PointFilter..." .  It means that after Ipopt
> has calculated x(k) at step k, I shall modify x(k) for the next step k+1.  If
> I did so, what should I mention to keep searching direction?  Is it possible
> in optimisation?  Could it make my problem becomes concave?

I don't understand your question.  If you have additional restrictions on
your variables, you should pose those as constraints and let the algorithm
handle them in a natural way.  Trying to change the iterates outside the
algorithm is most probably not going to work well.

Andreas