[SCIP] unresolved numerical troubles

[Vladimir V. Voloshinov]

Dear colleagues,
I found your discussion on
http://listserv.zib.de/pipermail/scip/2021-February/  because I began to
receive the same messages:
*(node xxxxxx) numerical troubles in LP yyyyyyy -- unresolved*
It was almost the first time I saw them in the SCIP log.
It was a global optimization problem with continuous variables, linear,
bilinear and SOS2 constraints.
E.g. the following dimension (but I saw that "numerical troubles" for
smaller problems of the same kind as well)



*presolved problem has 337 variables (0 bin, 0 int, 0 impl, 337 cont) and
81 constraints     20 constraints of type <SOS2>     40 constraints of type
<linear>     21 constraints of type <nonlinear>*

Can you, please, answer a couple of questions:
1. As far as I understood, Marc Pfetsch recommended SCIP build with IPOPT,
but my SCIP was built with it already.
*I use SCIP 8.0.1/Soplex 6.0.1/Ipopt 3.14.9 now.*
Is it possible to "stimulate" SCIP to use "more Ipopt" than LP by some SCIP
parameters ?

2. After the 10 warnings I see
*(node 30317) numerical troubles in LP 175340 -- unresolved -- further
messages will be suppressed (use display/verblevel=5 to see all)*
Although SCIP continued solving to the normal completion it seems that
these troubles slow down the solving process.
Is it true? Are there any known evaluations of loss in performance for the
case?

Sincerely yours,
Vladimir Voloshiov.

On Fri, Feb 26, 2021 at 2:00 PM James Cussens <james.cussens at bristol.ac.uk>
wrote:

> Hi Marc,
>
> Yes, using IPOPT is a good idea. I've used IPOPT before on similar
> instances (not ones with numerical problems) and, much to my surprise,
> solving was slower. But avoiding
> these numerical problems is more important and perhaps I can get to the
> bottom of (and hopefully avoid) the performance problem.
>
> Thanks,
>
> James
>
> James Cussens
> Dept of Computer Science, University of Bristol
> https://jcussens.github.io/
> Funded PhDs available in Bristol in the following areas: Data Science
> <http://www.bristol.ac.uk/cdt/compass/>, Interactive AI
> <http://www.bristol.ac.uk/cdt/interactive-ai/>, Cyber Security
> <http://www.bristol.ac.uk/cdt/cyber-security/> or Digital Health
> <http://www.bristol.ac.uk/cdt/digital-health/>.
> ------------------------------
> *From:* Scip <scip-bounces at zib.de> on behalf of Marc Pfetsch <
> pfetsch at mathematik.tu-darmstadt.de>
> *Sent:* 26 February 2021 10:51
> *To:* scip at zib.de <scip at zib.de>
> *Subject:* Re: [SCIP] unresolved numerical troubles
>
>
>
> Hi James,
>
> unresolved numerical troubles, means that all efforts of SCIP to get a
> feasible and optimal LP solution were fruitless. SCIP tries to tighten
> tolerances, changes the simplex method, restarts from scratch etc. in
> order to do so. If the LP solver fails to produce a solution, SCIP has
> to stop.
>
> In your particular case, there seems to be one simple thing to improve
> your situation: You should build SCIP with IPOPT - this solves your
> instance for me. IPOPT helps with finding feasible solutions, which
> seems to stablize the process.
>
> Best
>
> Marc
>
>
>
> On 26/02/2021 10.05, James Cussens wrote:
> > Hi all,
> >
> > I use a subMIQP for a pricer I have. Sometimes when I run it I get
> >
> > [solve.c:3912] ERROR: (node 1) unresolved numerical troubles in LP 32
> > cannot be dealt with
> >
> > I am keen to understand what might have led to such (fatal!) problems
> > and if I can avoid them. I attach an example of this problem. In the
> > attached file you have the original problem and the output when solving
> > the (presolved, transformed problem). The odd thing is that, in the
> > example I have attached, after presolving, the problem should not be too
> > hard to solve: all variables are continuous and we only have convex
> > constraints (22 second-order cone and 21 linear).
> >
> > One of my variables is fixed to a small positive value, perhaps that has
> > something to do with it.
> >
> > I am using SoPlex as the LP solver.
> >
> > James
> >
> > James Cussens
> > Dept of Computer Science, University of Bristol
> > https://jcussens.github.io/
> > Funded PhDs available in Bristol in the following areas: Data Science
> > <http://www.bristol.ac.uk/cdt/compass/>, Interactive AI
> > <http://www.bristol.ac.uk/cdt/interactive-ai/>, Cyber Security
> > <http://www.bristol.ac.uk/cdt/cyber-security/> or Digital Health
> > <http://www.bristol.ac.uk/cdt/digital-health/>.
> >
> > _______________________________________________
> > Scip mailing list
> > Scip at zib.de
> > https://listserv.zib.de/mailman/listinfo/scip
> >
> _______________________________________________
> Scip mailing list
> Scip at zib.de
> https://listserv.zib.de/mailman/listinfo/scip
> _______________________________________________
> Scip mailing list
> Scip at zib.de
> https://listserv.zib.de/mailman/listinfo/scip
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listserv.zib.de/pipermail/scip/attachments/20221125/64e8c47a/attachment.html>