[Csdp] how to make solver not give up

[Brian Borchers]

Jakub;

An exit code of 3 indicates that the solver was able to obtain a nearly
optimal solution that did not quite reach the desired tolerances for primal
infeasibility, dual infeasibility, or the duality gap.  If CSDP gets within
a factor of 100 of the desired tolerances but can't improve the solution
further, it will stop with this exit code.  If you're using the default
tolerances of 1.0e-8, then your solution should still be good to about six
digits- It may be that this solution is actually suitable for your
purposes.

CSDP gives up in these cases because it simply can't produce a more
accurate solution using its algorithm in double precision arithmetic.  This
can happen because the linear constraints are badly scaled or otherwise
badly conditioned or because of the interaction of the linear constraints
and the semidefiniteness constraint (CSDP actually requires solutions to be
positive definite and have Cholesky factorizations)- there may simply not
be any primal solution X which has a Cholesky factorization in double
precision and satisfies the linear constraints up to the full tolerance.
The same issue can occur with Z in the dual solution.

If you would be willing to send me your problem (in the form of a .dat-s
file), I'd be happy to take a look at it.


On Fri, Sep 7, 2018 at 3:50 AM Jakub Sliacan <jakub.sliacan at gmail.com>
wrote:

> Hi!
>
> I'm fairly new to SDP problems and I encountered a situation where I
> submit a job to a cluster (run CSDP on a big dat-s file) and while it's
> going fairly fast, the CSDP gives up very quickly (after only 80 iterations
> ~10mins of runtime). The return code is 3 (I can't replicate it now because
> cluster is down, but I think it's not 7). I was wondering if there is a way
> to make CSDP not finish with a near optimal solution but keep struggling
> for a return code 0 (say at least for maxiters=500, then return whatever).
>
> Thanks for any pointers!
> Jakub
>
> _______________________________________________
> Csdp mailing list
> Csdp at list.coin-or.org
> https://list.coin-or.org/mailman/listinfo/csdp
>


-- 
Brian Borchers                          borchers at nmt.edu
Department of Mathematics      http://www.nmt.edu/~borchers/
New Mexico Tech                       Phone: (575) 322-2592
Socorro, NM 87801                   FAX: (575) 835-5366
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.coin-or.org/pipermail/csdp/attachments/20180907/f0db49a8/attachment.html>