[Ipopt] Hessian Evaluations

[Andrew Spiteri]

Hi all,

I am using IPOPT 3.11.8 with MATLAB interface. I have managed to get the 
gradients, Jacobians and Hessians working with INTLAB. I've used IPOPT's 
derivative checker to make sure there are no issues in both the first 
order and second order derivatives.

However, I am seeing very different behaviour when using the 
'limited-memory' option for the Hessian evaluation instead of the 
'exact'. In fact, the 'limited-memory' option seems to be giving a 
considerably better performance (finding solutions in less iterations).
I've read a little on the 'limited-memory' option which says that it 
might give superior performance in the case of dense Hessians. While my 
Hessians are quite sparse, I am currently setting *hessianstructure* to 
*tril(ones(nvars))*, thus assuming all entries can be non-zero. Could 
this be the problem causing lesser performance of the 'exact' option? Or 
perhaps anyone has other suggestions? Thanks in advance.

Andrew Spiteri
Institute of Aerospace Technologies
University of Malta
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.coin-or.org/pipermail/ipopt/attachments/20170412/ece16b71/attachment.html>