[Coin-discuss] ClpModel, optimize for delayed constraint generation

[Sebastian Nowozin]

Hello everybody,

I am using OsiClpSolverInterface for delayed constraint generation,
using a sequence of resolve()/addRow() calls.  How to best optimize for
this setting?

So far, I do:
    OsiClpSolverInterface* si = new OsiClpSolverInterface;
    ClpSolve lp_options;
    lp_options.setPresolveType(ClpSolve::presolveOff);
    si->setSolveOptions(lp_options);

    // Initial static constraints
    CoinPackedMatrix* matrix = new CoinPackedMatrix(false, 0.5, 0);
    // ... (setup matrix, var*, row*
    si->loadProblem(*matrix, varLB, varUB, objective, rowLB, rowUB);

    int constraints_found = 0;
    do {
        si->resolve();  // Warm-start solving (we only add constraints)

        RemoveInactiveConstraints();  // delete non-binding constraints
        constraints_found = AddViolatedInequalities();    // add cuts
    } while (constraints_found > 0);

Here, RemoveInactiveConstraints uses si->deleteRows, and
AddViolatedInequalities() uses si->addRow() repeatedly.  From profiling,
I see that quite some time is spend in addRow().

I tried using:
    if (si->getModelPtr()->matrix()->isColOrdered()) {
        si->getModelPtr()->matrix()->setExtraGap(0.5);
    } else {
        si->getModelPtr()->matrix()->setExtraMajor(0.5);
    }

with no result.  Strangely, the si->getModelPtr()->matrix() is always
column ordered, even if my initial CoinPackedMatrix using in
si->loadProblem is row-ordered.  Calling reverseOrdering() on the matrix
leads to a crash in the solver, so it does not seem to be working.

How to optimize OsiClpSolverInterface for repeated calls of addRow() and
resolve()?  I do not know the exact number of cuts a-priori (around
5000-10000 cuts per iteration).

Thank you,
Sebastian