[Coin-ipopt] Re: need your comment

[Andreas Waechter]

Hi Rohallah,

> I deal with topology optimization (constrained PDE optimization) my PDE is
> non-linear heat transfer equation with phase change; aim is step forward to
> automation riser design in casting processes.
>
> I have several setup, in one of them number of design variables and
> constraints are high, 1-3 constraint per design variable, for this I select
> IPOPT as method for solution of optimality condition system.
>
> But in the other setup, I have large number of design variable and only one
> constraint. Is IPOPT for this good choice? Note that I can form augmented
> Lagrangian (with add constrain) and find its lagrange multiplier with inner
> iteration (especially as evaluation of constrain is very cheap in comparison
> with evaluation of objective function),
>
> e.g.
> calce d obj
> A = A0 (guess, A is multiplier)
> while constrain is not satisfied do
> dL = d obj + A d constrain  (d constrain  is constant value for my case)
> Xnew_hat = Xold + alfa * dL
> End wihle
> Xnew = Xnew_hat
>
> Step size alfa can be found with line search or CG method, especially I can
> use quadratic approximation for finding cg step size.
> What is your idea/comment?

I would think that Ipopt should do a good job at solving problems with 
only a few constraints.  If you have no inequality constraints of bounds 
(which you seem to imply), it is essentially the same as an SQP method.

I suggest you just give it a try.  If you dont want to code your problem 
in C/C++ before committing to Ipopt, you could just write an AMPL model 
and solve it from there (then you don't have to code derivative 
information).  Information on AMPL is here: http://www.ampl.com/ (you can 
get a free version for small problems there), and if you want to solve 
larger problems, you can use the NEOS Solvers:

http://www-neos.mcs.anl.gov/neos/solvers/index.html

I hope this helps,

Andreas