[SCIP] Problem avoiding Farkas Pricing

[Gerald Gamrath]

Dear Fernando,

Farkas pricing is called if and only if your restricted master LP is 
infeasible. The Farkas multipliers actually give you a proof for this, 
so you could verify that this is the case.

I don't know what could go wrong in your case. How do you add the 
artificial variables? You should use SCIPaddVar() not 
(SCIPaddPricedVar()) and mark them to be initial. In your PRICERFARKAS 
callback, you could write the current problem (SCIPwriteTransProblem()) 
and see if this problem looks like you expect it to look. It might be 
that some of the variables are not in the LP and need to be added to it 
by SCIP automatically. In that case, the delay flag of your pricer 
should be set to TRUE. And finally, domain propagation may fix your 
artificial variables to 0 because of their high objective coefficient. 
Please try to disable domain propagation by setting 
propagating/maxroundsroot and propagating/maxrounds to 0.

Best,
Gerald

Am 25.09.2015 um 18:12 schrieb Fernando Afonso:
> Hi there,
>
> I have a branch-and-cut-and-price algorithm implemented using SCIP. It 
> works well!
> But when a given node gets infeasible and Farkas pricing is called, 
> the performance of my pricing algorithm is really poor. For my 
> specific case, to assign costs to 0 at objective function to run 
> Farkas pricing, makes the dominance rules loosen.
>
> A given node gets infeasible due to branching constraints (I 
> implemented my own branching rule). Because of that, I tried to 
> include an artificial variable with a high cost for each branching 
> constraint I include in the model. But it's not working. 
> Theoretically, the model would be always feasible after to include 
> such artificial variables, but Farkas pricing is still called.
>
> Do you guys have an idea what should I do to avoid Farkas pricing to 
> be called?
>
> Thank you in advance,
>
> Fernando
>
>
> _______________________________________________
> Scip mailing list
> Scip at zib.de
> http://listserv.zib.de/mailman/listinfo/scip

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listserv.zib.de/pipermail/scip/attachments/20151006/b9aa4b76/attachment.html>