[Cbc] Intermediate results via CbcModel.bestSolution()

[John Forrest]

If you are going in through CbcMain then it is a little bit more
complicated.  I have put a crude fix into CbcSolver.cpp as it does not
affect Cbc library and can't introduce a bug.  This is the fix (stable)
-

After line 6237 of CbcSolver.cpp I put

  cbcPreProcessPointer = &process;

and at line 3753

CglPreProcess * cbcPreProcessPointer = NULL;

That saves a externally referenceable pointer to the CglPreProcess
object.  I said it was crude

Then I modified driver4.cpp to add

#include "CglPreProcess.hpp"
extern CglPreProcess * cbcPreProcessPointer;

and modified code in event handler (note original comment about
preprocessed model) to be -

     // If preprocessing was done solution will be to processed model
      int numberColumns = model_->getNumCols();
      const double * bestSolution = model_->bestSolution();
      assert (bestSolution);
      printf("value of solution is %g\n",model_->getObjValue());
      if (!cbcPreProcessPointer) {
	for (int i=0;i<numberColumns;i++) {
	  if (fabs(bestSolution[i])>1.0e-8)
	    printf("%d %g\n",i,bestSolution[i]);
	}
      } else {
	/* need a bit more work.
	   Values of variables in preprocess problem will be correct
	   To find others you could cut and paste code for postProcess
	   in CbcSolver.cpp
	   Simpler to fix known variables in original - that MIP will usually
	   solve in zero nodes
	*/
	const int * originalColumn = cbcPreProcessPointer->originalColumns();
	for (int i=0;i<numberColumns;i++) {
	  if (fabs(bestSolution[i])>1.0e-8)
	    printf("%d (original column %d) %g\n",i,
		   originalColumn[i],bestSolution[i]);
	}
      }
  
It seemed to work.  I hope that helps.

Haroldo - I hope that gives you enough information.

John Forrest
On Thu, 2010-07-15 at 16:50 -0400, acw at ascent.com wrote:
> The quoted conversation below suggests that up until recently we've
> simply been dodging bullets, submitting problems that don't happen to
> trigger compression; or perhaps in our recent upgrade some of the
> things we changed caused compression to occur where it hadn't up until
> now. 
> 
> Our present code is very simple, following the outlines of some of the
> example drivers in the Cbc/examples directory.  Once the model has
> been set up, we call CbcMain0 and then CbcMain1.  By the time we call
> CbcModel->solver()->getColSolution(), the solution has been
> decompressed and everything is good.  So there's no visible messing
> about with CglPreProcess, and at the moment we don't know how to do
> what Dr. Forrest suggests.  Any additional advice or code samples
> would be welcome. 
> 
> 
> From: 
> Haroldo Santos
> <haroldo.santos at gmail.com> 
> To: 
> John Forrest
> <john.forrest at fastercoin.com> 
> Cc: 
> "cbc at list.coin-or.org"
> <cbc at list.coin-or.org> 
> Date: 
> 07/15/2010 01:36 PM 
> Subject: 
> Re: [Cbc] Intermediate results via
> CbcModel.bestSolution()
> 
> 
> ______________________________________________________________________
> 
> 
> 
> Hi Forrest,
> 
> originalColumns() returns a primal solution ? indexes to map to the
> original columns ?
> 
> On Thu, Jul 15, 2010 at 1:39 PM, John Forrest
> <john.forrest at fastercoin.com> wrote:
> > bestSolution() never decompressed solution so you just must have
> been
> > lucky.
> >
> > If you did CglPreProcess and pass a pointer to that in to the
> handler
> > using bestSolution() then you can pick up originalColumns() and put
> the
> > solution back into the original model.
> >
> > John Forrest
> >
> >
> >
> > On Thu, 2010-07-15 at 11:27 -0400, acw at ascent.com wrote:
> >> In our application, we pose biggish problems to Cbc, which then
> goes
> >> off and spends quite a while solving them.  Even before we get the
> >> optimal solution, we can make use of any solutions Cbc finds along
> the
> >> way.
> >>
> >> So, during the solution process, when we see that Cbc has improved
> its
> >> best result, we query the CbcModel using bestSolution().
> >>
> >> Recently we've started encountering a problem where the returned
> >> solution is much shorter than expected.  Looking at the number of
> >> columns in the solution makes us suspect that we are seeing
> solutions
> >> to the processed model, and not to the original one.  That is, Cbc
> is
> >> boiling the problem down to a smaller, equivalent problem, and
> solving
> >> that; when it's completely finished it inverts the transformation
> to
> >> give a solution to the original problem.  But if we interrupt it,
> we
> >> find an intermediate solution to the transformed problem, not the
> >> original.
> >>
> >> This makes a certain amount of sense, but it didn't use to happen.
>  In
> >> a previous Cbc version, either the model was not compressed in the
> >> same way, or bestSolution() inverted the compression regardless of
> >> whether the solution was complete.
> >>
> >> Is this change in behavior intentional?  If so, does there remain
> any
> >> supported way to extract an intermediate solution to the problem as
> >> posed?
> >>
> >> Thank you; our experience with Cbc has been overwhelmingly
> positive,
> >> and we're sure there's an easy solution to this glitch as there
> have
> >> been for every problem we've encountered in the past.
> >> _______________________________________________
> >> Cbc mailing list
> >> Cbc at list.coin-or.org
> >> http://list.coin-or.org/mailman/listinfo/cbc
> >
> >
> > _______________________________________________
> > Cbc mailing list
> > Cbc at list.coin-or.org
> > http://list.coin-or.org/mailman/listinfo/cbc
> >
> 
> 
> 
> -- 
> =============================================================
> Haroldo Gambini Santos
> Computing Department - Universidade Federal de Ouro Preto - UFOP
> email: haroldo [at ] iceb.ufop.br
> home/research page: http://www.iceb.ufop.br/decom/prof/haroldo/
> 
> "Computer science is no more about computers than astronomy
> is about telescopes." Edsger Dijkstra
> 
> 
> _______________________________________________
> Cbc mailing list
> Cbc at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/cbc
> 
> _______________________________________________
> Cbc mailing list
> Cbc at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/cbc