[Clp-tickets] [COIN-OR Linear Programming Solver] #24: Clp QP simple simplex failure

[COIN-OR Linear Programming Solver]

#24: Clp QP simple simplex failure
------------------------+---------------------------------------------------
 Reporter:  nowozin     |       Owner:  somebody
     Type:  defect      |      Status:  new     
 Priority:  major       |   Milestone:          
Component:  component1  |     Version:          
 Keywords:              |  
------------------------+---------------------------------------------------
 Using Clp trunk from 2008/09/15, I setup a QP problem as follows:
 {{{
     ClpSimplex* simplex = new ClpSimplex();
     CoinPackedMatrix* matrix = new CoinPackedMatrix(false, 0, 0);
     // ...
     matrix->setDimensions(0, number_columns);
     // ...
     simplex->loadProblem(*matrix, varLB, varUB, objective, rowLB, rowUB);
     // Setup the quadratic part of a simple "x'Ix + f'x" objective
     // ...
     simplex->loadQuadraticObjective(number_columns, start, column, w_q);
     // Add some rows (~2000) of the form
     for (...) {
         // ...
         simplex->addRow(dim + 1, col_index, col_el, lb, COIN_DBL_MAX);
     }
     // Solve
     simplex->primal();
 }}}

 Clp produces the following output which seems to give the correct
 objective
 2310:
 {{{
 Clp0024I Matrix will be packed to eliminate 307230 small elements
 Clp0006I 0  Obj 0 Primal inf 2310 (2310) Dual inf 2310 (2310)
 Clp0029I End of values pass after 0 iterations
 Clp0006I 0  Obj 0 Primal inf 2310 (2310) Dual inf 2310 (2310)
 Clp0006I 200  Obj 0 Primal inf 2110 (2110) Dual inf 2110 (2110)
 Clp0006I 400  Obj 0 Primal inf 1910 (1910) Dual inf 1910 (1910)
 Clp0006I 600  Obj 0 Primal inf 1710 (1710) Dual inf 1710 (1710)
 Clp0006I 800  Obj 0 Primal inf 1510 (1510) Dual inf 1510 (1510)
 Clp0006I 1000  Obj 0 Primal inf 1310 (1310) Dual inf 1310 (1310)
 Clp0006I 1200  Obj 0 Primal inf 1110 (1110) Dual inf 1110 (1110)
 Clp0006I 1400  Obj 0 Primal inf 910 (910) Dual inf 910 (910)
 Clp0006I 1600  Obj 0 Primal inf 710 (710) Dual inf 710 (710)
 Clp0006I 1800  Obj 0 Primal inf 510 (510) Dual inf 510 (510)
 Clp0006I 2000  Obj 0 Primal inf 310 (310) Dual inf 310 (310)
 Clp0006I 2200  Obj 0 Primal inf 110 (110) Dual inf 110 (110)
 Clp0006I 2310  Obj 0
 Clp0000I Optimal - objective value 0
 Clp0006I 0  Obj 2310
 Clp0000I Optimal - objective value 2310
 }}}

 However, I check, as I do for LP's, the following
 {{{
     if (simplex->isProvenOptimal() == false) {
         // error ...
     }
 }}}
 which is triggered and dumps:

 {{{
 Quadratic Program Solver failed.
 Problem: 2443 variables, 2310 rows.
 Written problem to file "QP-CRASH.qps" for analysis.
          problem sense: 1
 STATUS:  numerical difficulties: no
 STATUS:       primal infeasible: YES
 STATUS:         dual infeasible: no
 STATUS: iteration limit reached: no
 }}}

 If I ignore the isProvenOptimal() check and try to obtain the objective
 using
 {{{
     const double* primal = simplex->primalColumnSolution();
 }}}
 then it is all-zero, which is not a feasible point.

 I checked the dumped QPS file and solved the problem using BPMPD which
 solves
 it with objective 2310 (see attached output logfile).

 Am I using the interface incorrectly or is this a bug in Clp?
 Also, if I want to disable any kind of presolve, is there a way to do this
 using the ClpSimplex and ClpModel interfaces?  (In initialSolve I can pass
 a ClpSolve object, but how to do this just by calling primal()?)

 Thanks.

-- 
Ticket URL: <https://projects.coin-or.org/Clp/ticket/24>
COIN-OR Linear Programming Solver <http://projects.coin-or.org/Clp>
A linear programming solver.