[Coin-ipopt] 2 small updates

[Andreas Waechter]

Hi,

I just made two minor modification to the code in COIN; they are already
available in the CVS repositiory, and will also be reflected in tomorrow's
tarball...

1. I corrected the rule for the superlinear decrease of the barrier
parameter (thanks to Dominique Orban and Richard Waltz for telling me
about it!).  This new version might now truly lead to fast local
convergence (if second derivatives are used); you might safe a few
iterations at the end.

2. After Ned's message from December I had a look at the computation time
for the limited-memory quasi-Newton updates.  I'm not able to find
anything to make the code work faster (I probably was a little
overoptimistic recommending this quasi-Newton option as the new holy
grail...).  However, I decreased the default value of the parameter ILMLEN
from 20 to 6 which should decrease the CPU time per iteration (but might
lead to a larger number of overall iterations and possibly less
robustness).  I hope that this offers a reasonable compromise between
speed and robustness.  In any case, if you are using the limited-memory
BFGS option (IQUASI=6 or -6), it might be worthwhile playing with ILMLEN.

Cheers,

Andreas