[Ipopt] Feasible iterations only - again

[Horand Gassmann]

Fabian Wein <fabian.wein at lse.e-technik.uni-erlangen.de> wrote:

> In my current model, the objective is only defined for a feasible  
> design. Hence the nonlinear inequalitiy constraints
> need to be feasible for the nonconvex nonlinear problem.
>
> If found an old thread: "Feasible iterations only"  
> http://list.coin-or.org/pipermail/ipopt/2007-May/000778.html
> Here Andreas states that this is not in Ipopt and not planned.
>
> I have not much hope, but anyway - has this changed meanwhile?  
> Anyone knows a solver with this feature? E.g. SNOPT, KNITRO?

I asked Andreas a related question recently. His answer, which made a  
lot of sense to me, was to simply use a zero function as the  
objective. For your problem this would mean that you could solve the  
feasibility problem first, then use the known feasible point as a  
starting point for the optimization phase.

But if you have nonconvex functions, then Ipopt may not be the best  
solver, anyway. Among the COIN-OR solvers you may want to try Couenne,  
and you can work within OS (Optimization Services  
https://projects.coin-or.org/OS) to let you switch seamlessly between  
the two.

HTH

gus

-------------------------------------------------------

Horand I. Gassmann, Professor

School of Business Administration, Dalhousie University
6100 University Avenue, PO Box 15000
Halifax, Nova Scotia, Canada, B3H 4R2
ph. (902) 494-1844
fax (902) 494-1107

http://myweb.dal.ca/gassmann/