[Scip] much time when solving LP in B&price.

[Cristina Núñez del Toro]

Hello all,

I'm have implemented a Branch&Price algorithm but I have some questions
about the LP optimization behavior. When I try a quite big instance, my
implementation spends a lot of time obtaining the LP solution (not only for
the pricing routine, after adding variables but also when the first time in
the node). So, before thinking about deleting variables and/or changing
parameters for remove columns, I would like to have clear first some issues
about the LP re-optimization. For example:

1) How does LP is solved after adding new (pricing) variables? it is used
primal simplex, dual simplex or it is optimized from scratch?

2) How does the LP is solved when it first enters to a node after
branching? it is used primal simplex, dual simplex or it is optimized from
scratch?

About 1), my desire would be to re-optimize using primal simplex, so the
new variables could be the very first candidates to enter to the basis. In
particular, I would like to try 'partial' LP pricing strategy as an option
in reducing the cpu time. Changing this parameter is applied only if I use
SoPlex as LP solver or it is applied to any other LP solver (Cplex for
example) as well?

Another thing that could be related (or not) with this is:

3) How does the variable's pricing score in "SCIPaddPricedVar( SCIP * scip,
SCIP_VAR* var, SCIP_Real score)" is related with the LP reoptimization. I
really appreciate if you could give me more information about this.














Best regards,


-- 
---
Cristina Nuñez
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listserv.zib.de/pipermail/scip/attachments/20150421/16b072df/attachment.html>