[Ipopt] Exact hessian with significantly slower convergence than limited memory

[Roberto Verdelli]

I'm solving this problem with about 200 variables.
If I use the limited-memory hessian approximation I can get a solution after about 1 minute of computation.If I provide IpOpt the exact hessian the convergence slows down a lot and I get a solution after 15 minutes or more.
The starting point is infeasible and very far from the solution. The solution is something about 0.6, both runs start with objective = 300 but when I use the limited memory approximation I can see the  objective decreasing quickly to 1 and then converging to 0.6 at 'normal' speed. Whenever I provide the exact hessian I get a different behaviour: the objective starts at 300 and decreases at what seems a constant rate until the solution is reached.
The exact hessian is calculated correctly: I fill only the lower triangle and the derivative checker doesn't show any error.
I expected faster convergence with the exact hessian, why this behaviour?


 		 	   		  
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.coin-or.org/pipermail/ipopt/attachments/20140206/05dfdf13/attachment.html>