[Scip] Fwd: Fwd: getting column coefficients and realloc error

[Mattia Barbieri]

> Hi Mattia,
>
> >  - In my problem, I am doing branch on arcs. I edited the branch plugin
> > (like explained in documentation), and now I select the "longest" arc of
> >  the most fractional variable (a path). The best way to lead with my
> >  problem would be to select an arc providing the best lowerbound. To
> doing
> >  so, I need to loop over all arcs of current fractional variables, and
> try
> >  to solve the LP problem with each of these arcs. I think this is a sort
> of
> >  DIVING approach. But how can I do this in practice? Should I create a
> new
> >  Primal Heuristic plugin that explore one level of the subtrees
> associated
> >  whit current active (leaf) nodes?
> You can use the diving mode in SCIP. With the method SCIPstartDive() SCIP
> changes it stage into diving. With the methods
> - SCIPchgVarObjDive()
> - SCIPchgVarLbDive()
> - SCIPchgVarUbDive()
> - SCIPgetVarObjDive()
> - SCIPgetVarLbDive()
> - SCIPgetVarUbDive()
> - SCIPsolveDiveLP()
> - SCIPgetLastDivenode()
> you can control the dive (see scip.c) and with SCIPendDive() you reset SCIP
> into normal mode. An example for such a diving can be found in one of the
> diving heuristics such as "heur_coefdiving.c". The diving within SCIP is
> not
> restricted to heuristics it can also be performed in your brachching
> Plugins.
>

In my branching plugin, I try doing this:

 for (int v = 0; v < nlpcands; v++) {

        cVar = lpcands[v];

        *//include var*
        SCIPstartDive(scip);

        SCIPchgVarLbDive(scip, cVar, 1.0);
        SCIPchgVarUbDive(scip, cVar, 1.0);
        SCIPsolveDiveLP(scip, -1, error);

        LPval1 = SCIPgetLowerbound(scip);

        SCIPendDive(scip);

       * //exclude var*
        SCIPstartDive(scip);

        SCIPchgVarLbDive(scip, cVar, 0.0);
        SCIPchgVarUbDive(scip, cVar, 0.0);
        SCIPsolveDiveLP(scip, -1, error);

        LPval1 = SCIPgetLowerbound(scip);

        SCIPendDive(scip);

        LPval = MIN(LPval1, LPval2);

        *//take max of LPval of all fractional variables*
        if (LPval > bestLPval) {
            bestLPval = LPval;
            bestVarIndex = v;
        }
    }

But what I get is an incorrect primal bound (integer) solution. I suppose
this happens because a change in Lower/Upper bound in this way is done
permanently, while what I hope to get is a change only during the time I
solve the LP to select the best variable. Then I want the variable's
Upper/Lower bound come back to its original values, letting me going on to
the usual branching process without any changes in the variables (I will
include or exclude "arcs" from paths, not entire variables).

> -  After few nodes the actives variables are a lot (1000 to 20k variables,
> > considering 70-nodes problems). These are all the variables that my
> pricer
> > plugin find, but in practice only a half (or less) are active columns of
> >  the current LP at one node (I suppose that SCIP has a own mechanism to
> >  define a pool of active columns).
> Yes that is the case.
>
> > Thousands of columns are anyway too
> >  much, because the LP solving process is very slow. Is there anyway to
> >  better manage this problem?
> I am not sure if I got your question right. I try to answer it in the
> follwing
> way:
> There is one problem in the current version of SCIP. It is not possible to
> remove variables from the problem. In your case I would suggest to select
> only
> a certain number of the variables your pricer creates for adding to your
> problem in each round. That  might help to reduce the amount of useless
> variables in the end.
>

Instead it may help to remove from the current LP variables whit a very high
positive (min object sense) reduced cost. How can I deal whit this approach?

>
> > - A more tecnhical question. In relation whit the previous question, I
> >  think in my problem would help a faster LP solver, like Gurobi. I try to
> >  link it whit SCIP, but I get a lot of error in compilation such
> undefined
> >  funcionts (I read in some questions of other users that Gurobi is not
> >  fully supported at the moment). I installed the version 3.00 of Gurobi.
> In case of Gurobi the problem comes from the fact that Gurobi does not
> support
> all callback methods which are needed be integrated into a MIP solver like
> SCIP. This will probably change over the next versions of Gurobi and we
> will
> update the SCIP Gurobi interface accordingly.
>
> > I try also to link SCIP whit CPLEX 11.1: the license is for the x86
> version,
> >  but I have to compile SCIP in x86_64 mode because the underlying have an
> >  X86_64 operating system. If I try to force the installation of SCIP whit
> >  parameter ARCH=x86 I get this error
> Since CPLEX is now free for academic people you could get a CPLEX x86_64
> license. Otherwise try the command suggested by Stefan V.
>

Ok I will try this way.

>
> -- Stefan
>

Thanks in advance.


-- 
Mattia Barbieri
barbieri.mattia at gmail.com



-- 
Mattia Barbieri
barbieri.mattia at gmail.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://listserv.zib.de/mailman/private/scip/attachments/20100519/a7371595/attachment.html