[Coin-ipopt] max constraint violation -- user parameter question

[Andreas Waechter]

Hi Berry,

Many thanks for posting your question to the Ipopt mailing list.

> I have a model that ipopt solves for small datasets.  For
> larger datasets, however, it ends with error 10 after many
> thousands of iterations.

Hmmm..  That's too bad :)

> During the course of the computation, I get messages such as
> the following two examples (from the same model with different
> settings to model-specific parameters).
>
> resto_filter: It is possible that filter blocks progress.
>        Decrease max allowed constraint violation from   0.323379295
>         to   0.0323379295 and delete current filter.
>
> resto_filter: It is possible that filter blocks progress.
>        Decrease max allowed constraint violation from   5.47036108E+10
>         to   5.47036108E+09 and delete current filter.
>
> Do these message represent suggestions that I
> change IITERBLOCKMAX and/or DWATCHTOL (or some other
> parameter), or are they ipopt reports about actions it
> is taking on its own?

Sorry for the confusion.  That message is just information - it means that
Ipopt is decreaseing the maximal allows constraint violation..., not that
the user has to take some action.  (I meant to clarify this output
message, must have forgotten to do that - will do so now).

> Full disclosure:  I have the
> ipopt paper, but haven't read it, so I don't know the
> algorithm.

No problem - at least you have the paper :-)  The message relates to the
heuristic "Case I" on page 11 of the paper.  I guess you already made that
connection because you found the parameters IITERBLOCKMAX and DWATCHTOL.

However, changing those parameters probably doesn't make a big difference.
It is difficult to make suggestions without more information.  The exit
code 10 could mean that the constraints in your problem formulation are
almost linearly dependent, or maybe that some constraints or variables are
not well scaled.  If you want, you can send me the IPOPT.OUT file for
printlevel 3 (in that case directly to andreasw at watson.ibm.com, so that we
don't send a huge message to the mailing list), and I can try to see if
from that I can get some idea of what might be changed.

Cheers,

Andreas