[Scip] unexplainable behaviour of probing

[Tobias Achterberg]

Hi Frank,

to give a comprehensive answer I had to take a closer look at the source code, but to make
it short: probing is (in its current implementation) not allowed to be called in a pricer.

To give you some clue on what is happening:

At the end of probing, the LP is reset and solved again to reproduce the solution status
in the LP solver. This is done in the method SCIPtreeEndProbing() in src/scip/tree.c. And
this method also updates the lower bound of the current node using the LP objective value.
But because the LP objective value is not a valid lower bound in the presence of a pricer,
I guess that this gives you a wrong result.

You could try and just remove the line

            SCIP_CALL( SCIPnodeUpdateLowerboundLP(tree->focusnode, set, stat, lp) );

from the SCIPtreeEndProbing() code. Actually, I don't think that this line is needed
anyway, but the final decision has to be made by the SCIP developers. But please give it a
try and see what happens...


Tobias


Frank Hennig wrote:
> Dear SCIP experts,
> 
>  
> 
> I want to use probing during a pricer. If I do that the optimization log
> behaves strange.
> 
>  
> 
> If I disregard my probing and just call SCIPstartProbing followed by
> SCIPnedProbing during my pricer it, at some point, sets the dual bound
> to the known primal bound with a resulting gap of 0%.
> 
>  
> 
> The next log line reports a heuristically found solution that improved
> the primal solution and reports the new primal (upper) bound and the
> current dual (lower) bound. The lower bound is larger than the upper bound.
> 
> The final log line corrects the dual bound to the lower primal bound
> value and concludes optimality.
> 
>  
> 
>  
> 
> The whole thing doesn’t happen without the two probing code lines.
> 
>  
> 
> How can this be?
> 
>  
> 
>  
> 
> Regards,
> 
>  
> 
> Frank
> 
>  
> 
>  
> 
> 
> ------------------------------------------------------------------------
> 
> _______________________________________________
> Scip mailing list
> Scip at zib.de
> http://listserv.zib.de/mailman/listinfo/scip