[Scip] MISPRICE IN SMOOTHING STABILIZATION
Gerald Gamrath
gamrath at zib.de
Fri Apr 24 10:49:19 CEST 2015
Dear Diego,
you are right, you need to handle the case of a misprice explicitly. If
your pricer finishes without adding variables (and your result pointer
is set to SUCCESS), SCIP will assume that there are no improving
variables and use the current LP value as a dual bound.
This means, whenever you are not sure if there are more improving
variables, you should set the result pointer to DIDNOTRUN and SCIP won't
use the LP value as dual bound. On the other hand, it will stop pricing
and perform branching, which you do not want to do either.
Therefore, your pricer should just handle this case. If it performed
pricing with smoothed dual values and does not find an improving
variable, it should just perform another round of pricing (without
returning from the PRICERREDCOST callback). In this round, you either
adjust your alpha to give a higher weight to the original dual values or
just use those alone and perform exact pricing without smoothed duals.
Best,
Gerald
On 21.04.2015 12:50, dponce at us.es wrote:
>
> Hello list,
>
> as part of a Branch-and-Price procedure, I'd like to apply
> stabilization in order to add less variables at the end to check
> optimality. As I readed in another thread, smoothing approach (Pessoa
> et al.) is a good option in SCIP.
>
> I have it implemented, but it avoids correct solutions. I think the
> problem is the following: when a misprice (there are columns to add
> but the pricing do not provide such columns using the convex
> combination) happens, the pricing detect that no variables were added
> and SCIP go out the pricing before time. How could I deal whith this
> problem?
>
> Thanks in advance.
>
> Diego.
>
>
>
> _______________________________________________
> 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/20150424/7297e431/attachment.html>
More information about the Scip
mailing list