[Coin-lpsolver] Precision for Clp solution

[John J Forrest]

Francois,

As you say, the formulation has some interesting numbers so in order to
solve it Clp scales the problem (I have tried in many ways and for some of
them Clp does not enjoy solving the problem and I may try and improve
performance) and while the scaled reduced costs are within tolerances the
unscaled ones are not.  This is a documented (well as much as anything in
Clp is documented) feature and Clp returns a secondary status of 3 which
means scaled problem optimal, unscaled has dual infeasibilities.  If you
get this status you can then set hints for OsiClp to do primal and do not
scale and then resolve.  This seems to clear up dual infeasibilities.

John


                                                                           
             Francois Margot                                               
             <fmargot at andrew.c                                             
             mu.edu>                                                    To 
             Sent by:                  coin-lpsolver at list.coin-or.org      
             coin-lpsolver-bou                                          cc 
             nces at list.coin-or                                             
             .org                                                  Subject 
                                       [Coin-lpsolver] Precision for Clp   
                                       solution                            
             04/10/06 10:19 AM                                             
                                                                           
                                                                           
                                                                           
                                                                           
                                                                           





What precision can we expect for the results obtained from Clp, using
Lapack and Blas libraries on a Linux machine? Is 10e-3 too much to
expect? (I just updated my copy of COIN this morning.)

I am solving an LP (admittedly with a numerically suspect formulation,
with some coefficients smaller than 10e-10) using:

   OsiClpSolverInterface *clp = new OsiClpSolverInterface;
   clp->readMps("bc1.mps");
   clp->initialSolve();

   const double *lb = clp->getColLower();
   const double *ub = clp->getColUpper();
   const double *rc = clp->getReducedCost();

   int ncol = clp->getNumCols();
   int nrow = clp->getNumRows();

   for(int i=0; i<ncol; i++) {
     if((rc[i] > 1e-4) && (ub[i] - lb[i] > 1e-4) && (x[i] > lb[i] + 1e-4))
{
       printf("### ERROR: var: %d  rc: %f   lb: %f  ub: %f  x: %f\n",
                  i, rc[i], lb[i], ub[i], x[i]);
       exit(1);
     }
     if((rc[i] < -1e-4) && (ub[i] - lb[i] > 1e-4) && (x[i] < ub[i] - 1e-4))
{
       printf("### ERROR: var: %d  rc: %f   lb: %f  ub: %f  x: %f\n",
                  i, rc[i], lb[i], ub[i], x[i]);
       exit(1);
     }
   }


I get:

### ERROR: var: 8  rc: -0.002807   lb: 0.000000  ub: 10.000000  x: 0.000000

Regardless of the initial formulation, the precision of the result seems
to be at most 10e-2. The mps file bc1.mps is available from

http://plato.asu.edu/ftp/milp/

Francois

_______________________________________________
Coin-lpsolver mailing list
Coin-lpsolver at list.coin-or.org
http://list.coin-or.org/mailman/listinfo/coin-lpsolver