[Coin-ipopt] Re: IPOPT observations

[Andreas Waechter]

Hi Ned,

> On some of my runs, IPOPT returns with IERR = 823. I find this unusual. I
> believe IERR should not be 823.
> Am I right?

This error message occurs if some dense matrix which has to be factorized
for the limit-memory BFGS option becomes singular.  Since the L-BFGS
option is very new, I wasn't sure when/if this could happen, and what
should be done in that case.  It might be helpful if you could send me the
IPOPT.OUT output file (generated with IPRINT = 10 or at least 3) [offline,
not to the mailing list] and I could have a look at it.

> I tried  ifull = 1 and  iquasi = 6. With these options, the execution is
> substantially slower than with  with ifull = 0 and iquasi = -3.
> I am puzzled by this.

Interesting you say this - someone else told me that he is also concerned
about the runtimes.  One way to reduce the CPU time for iquasi = 6 is to
choose a smaller value for ilmlen; this is the "length of the memory"
(number of (s,y) pairs) for the limited memory BFGS version (default is
20), but this might lead to overall more iterations.  I noted that the CPU
time per iteration increases quite a bit at the beginning, when more and
more of those (s,y) pairs are collected (up to the maximum number
specified by ilmlen), and I also find this somewhat surprising since for
each of those pairs only an additional backsolve of the linear system is
required and should not have such a noticable effect.  If I find some time
over the Holidays, I can see if I can hunt this down.

The best choice of parameters always depends on the problem
characteristics (number of variables, number of constraints, for the
reduced space option: how easy it is for the algorithm to find a good
basis or if it can even be provided by the user)...

By the way, I'm surprised you chose iquasi = -3, and not -2 or -1.  Did -3
work better for you?

Anyway, Merry Christmas for now,

Andreas