[Ipopt] unknown sparsity pattern; #non-zeros in constr. Jac. not allowed to change; objective and infeasibility grow despite filter

Martin Neuenhofen martinneuenhofen at googlemail.com
Fri Jan 20 07:05:00 EST 2017

Dear all,

I would like to add an attendum to question 3.

For that observed situation the initial guess was feasible at 10^-3. The
box constraints had been moved to cL=min([1.5*c(x0), 0.5 * c(x0), cL],[],2)
and analogously for cR, thus the x0 is far away from any inequality
borders. Ipopt evolves x and increases the feasibility and the
objective with each iteration.
This should not arise from the adaptive mu since x is not near to
inequality feasibility bounds. Are there potential explanations for this

Kind regards

2017-01-20 11:46 GMT+00:00 Martin Neuenhofen <
martinneuenhofen at googlemail.com>:

> Hi all,
> I have three questions [1,2,3].
> For my application I don't know the sparsitiy pattern of my Jacobian in
> advance. [1] Why does Ipopt want to know it in advance at all? Is a
> graph-based fill-in reducing reordering only computed once in the beginning
> or for what benefit?
> Calling my model with NaNs gives me by far more nonzeros than would
> practically occur in that Jacobian for any call with real values.
> Further, I have the bad feeling that when in Matlab my matrix has by
> properties of my current iterate another number of non-zeros then Ipopt
> will rearrange them falsely (by using the same index arrays for the sparse
> storage), will thus use a wrong Jacobian and consequently fail to converge
> towards feasibility. [2] Is this true?
> Why is it possible at all that Ipopt performs iterations where both
> objective and primal infeasibility grow? [3] Shouldn't the filter enforce
> that never one grows and always one strictly decreases (or in case any of
> these seems impossible that the solver exists)?
> Thanks for your answers.
> Kind regards,
> Martin
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.coin-or.org/pipermail/ipopt/attachments/20170120/0a5eb7fc/attachment.html>

More information about the Ipopt mailing list