[Ipopt-tickets] Re: [Ipopt] #45: Suggestion

[Ipopt]

#45: Suggestion
--------------------------------+-------------------------------------------
  Reporter:  kapurash           |       Owner:  andreasw
      Type:  enhancement        |      Status:  assigned
  Priority:  normal             |   Component:  Ipopt   
   Version:  3.3 (C++ Version)  |    Severity:  normal  
Resolution:                     |    Keywords:          
--------------------------------+-------------------------------------------
Changes (by andreasw):

  * status:  new => assigned
  * owner:  ipopt-team => andreasw

Comment:

 Hi,

 Yes, in principle that would be a great thing.  However, currently, Ipopt
 uses direct factorizations of the iteration matrix (see equation (13) in
 http://www.research.ibm.com/people/a/andreasw/papers/ipopt.pdf) and
 requires the full Hessian matrix.  (The limited-memory version also works
 with a similar matrix, but the Hessian in the linear system here is a
 diagonal matrix)

 One possible approach here is to use iterative linear solvers, which would
 require only products for the matrix, but this has not been implemented.
 We are working on integrating an iterative linear solver for the iteration
 matrix, but at this point, also this solver will require the entire
 Hessian matrix at once, since the preconditioner is computed with that
 information.

 Alternatively, the step computation could be done similar to how it is
 computed in KNITRO (which you might want to consider, since it probably
 doesn't need the entire Hessian - but KNITRO is not free), in a two-phase
 approach.  It uses a mix between direct and indirect linear solvers...
 Again, this however would not be a trivial change.  In case you are
 serious about helping, how much do you know about C++ and optimization
 methods...? ;-)

 Andreas

-- 
Ticket URL: <https://projects.coin-or.org/Ipopt/ticket/45#comment:1>
Ipopt <http://projects.coin-or.org/Ipopt>
Interior-point optimizer for nonlinear programs.