[Ipopt] Using Ipopt for "one-shot" optimisation

[Jens-Dominik Mueller]

Hello,

we are looking for an opt. package that we can couple to our flow and 
adjoint (gradient) solver for CFD. To reduce the heavy computational 
cost of the function and gradient evaluations, we are using a one-shot 
method that solves the KKT in a block coupled fashion as shown below.

In particular, after an initial convergence to get a rough approximation 
of functional and gradient, in the design loop we are continuously 
converging the state and adjoint while updating the design/control 
variables x. This is nicely done with a callback implementation of the 
optimiser, such as the classic Northwestern LBFGS that we currently are 
using.

Can we achieve sth. similar with Ipopt?

Many thanks, Jens.


- Initial convergence of flow and adjoint, say 1-2 O of magnitude
- evaluate functional and gradient

- design loop:
   call optimiser
   if ( optimiser says new_x )
     apply new_x to grid
     converge flow, adj some more
     evaluate functional and gradient
   else ( optimiser says done )
     done
   else ...
    ..
   end if


---------------------------------------------------------------
   Dr. Jens-Dominik Müller
   School of Engineering and Materials Science
   Mile End Road
   Queen Mary, University of London
   London, E1 4NS

   tel: (+44) (0) 20 7882 5421, fax: (+44) (0) 20 8983 1007