int CbcHeuristicDive::reducedCostFix (OsiSolverInterface* solver)
{
//return 0; // temp
#ifndef JJF_ONE
if (!model_->solverCharacteristics()->reducedCostsAccurate())
return 0; //NLP
#endif
double cutoff = model_->getCutoff() ;
if (cutoff > 1.0e20)
return 0;
#ifdef DIVE_DEBUG
std::cout << "cutoff = " << cutoff << std::endl;
#endif
double direction = solver->getObjSense() ;
double gap = cutoff - solver->getObjValue() * direction ;
gap *= 0.5; // Fix more
double tolerance;
solver->getDblParam(OsiDualTolerance, tolerance) ;
if (gap <= 0.0)
gap = tolerance; //return 0;
gap += 100.0 * tolerance;
double integerTolerance = model_->getDblParam(CbcModel::CbcIntegerTolerance);
const double *lower = solver->getColLower() ;
const double *upper = solver->getColUpper() ;
const double *solution = solver->getColSolution() ;
const double *reducedCost = solver->getReducedCost() ;
int numberIntegers = model_->numberIntegers();
const int * integerVariable = model_->integerVariable();
int numberFixed = 0 ;
# ifdef COIN_HAS_CLP
OsiClpSolverInterface * clpSolver
= dynamic_cast<OsiClpSolverInterface *> (solver);
ClpSimplex * clpSimplex = NULL;
if (clpSolver)
clpSimplex = clpSolver->getModelPtr();
# endif
for (int i = 0 ; i < numberIntegers ; i++) {
int iColumn = integerVariable[i] ;
double djValue = direction * reducedCost[iColumn] ;
if (upper[iColumn] - lower[iColumn] > integerTolerance) {
if (solution[iColumn] < lower[iColumn] + integerTolerance && djValue > gap) {
#ifdef COIN_HAS_CLP
// may just have been fixed before
if (clpSimplex) {
if (clpSimplex->getColumnStatus(iColumn) == ClpSimplex::basic) {
#ifdef COIN_DEVELOP
printf("DJfix %d has status of %d, dj of %g gap %g, bounds %g %g\n",
iColumn, clpSimplex->getColumnStatus(iColumn),
djValue, gap, lower[iColumn], upper[iColumn]);
#endif
} else {
assert(clpSimplex->getColumnStatus(iColumn) == ClpSimplex::atLowerBound ||
clpSimplex->getColumnStatus(iColumn) == ClpSimplex::isFixed);
}
}
#endif
solver->setColUpper(iColumn, lower[iColumn]) ;
numberFixed++ ;
} else if (solution[iColumn] > upper[iColumn] - integerTolerance && -djValue > gap) {
#ifdef COIN_HAS_CLP
// may just have been fixed before
if (clpSimplex) {
if (clpSimplex->getColumnStatus(iColumn) == ClpSimplex::basic) {
#ifdef COIN_DEVELOP
printf("DJfix %d has status of %d, dj of %g gap %g, bounds %g %g\n",
iColumn, clpSimplex->getColumnStatus(iColumn),
djValue, gap, lower[iColumn], upper[iColumn]);
#endif
} else {
assert(clpSimplex->getColumnStatus(iColumn) == ClpSimplex::atUpperBound ||
clpSimplex->getColumnStatus(iColumn) == ClpSimplex::isFixed);
}
}
#endif
solver->setColLower(iColumn, upper[iColumn]) ;
numberFixed++ ;
}
}
}
return numberFixed;
}
int main(int argc, const char *argv[])
{
/* Read quadratic model in two stages to test loadQuadraticObjective.
And is also possible to just read into ClpSimplex/Interior which sets it all up in one go.
But this is only if it is in QUADOBJ format.
If no arguments does share2qp using ClpInterior (also creates quad.mps which is in QUADOBJ format)
If one argument uses simplex e.g. testit quad.mps
If > one uses barrier via ClpSimplex input and then ClpInterior borrow
*/
if (argc < 2) {
CoinMpsIO m;
#if defined(SAMPLEDIR)
int status = m.readMps(SAMPLEDIR "/share2qp", "mps");
#else
fprintf(stderr, "Do not know where to find sample MPS files.\n");
exit(1);
#endif
if (status) {
printf("errors on input\n");
exit(77);
}
ClpInterior model;
model.loadProblem(*m.getMatrixByCol(), m.getColLower(), m.getColUpper(),
m.getObjCoefficients(),
m.getRowLower(), m.getRowUpper());
// get quadratic part
int * start = NULL;
int * column = NULL;
double * element = NULL;
m.readQuadraticMps(NULL, start, column, element, 2);
int j;
for (j = 0; j < 79; j++) {
if (start[j] < start[j+1]) {
int i;
printf("Column %d ", j);
for (i = start[j]; i < start[j+1]; i++) {
printf("( %d, %g) ", column[i], element[i]);
}
printf("\n");
}
}
model.loadQuadraticObjective(model.numberColumns(), start, column, element);
// share2qp is in old style qp format - convert to new so other options can use
model.writeMps("quad.mps");
ClpCholeskyBase * cholesky = new ClpCholeskyBase();
cholesky->setKKT(true);
model.setCholesky(cholesky);
model.primalDual();
double *primal;
double *dual;
primal = model.primalColumnSolution();
dual = model.dualRowSolution();
int i;
int numberColumns = model.numberColumns();
int numberRows = model.numberRows();
for (i = 0; i < numberColumns; i++) {
if (fabs(primal[i]) > 1.0e-8)
printf("%d primal %g\n", i, primal[i]);
}
for (i = 0; i < numberRows; i++) {
if (fabs(dual[i]) > 1.0e-8)
printf("%d dual %g\n", i, dual[i]);
}
} else {
// Could read into ClpInterior
ClpSimplex model;
if (model.readMps(argv[1])) {
printf("errors on input\n");
exit(77);
}
model.writeMps("quad");
if (argc < 3) {
// simplex - just primal as dual does not work
// also I need to fix scaling of duals on output
// (Was okay in first place - can't mix and match scaling techniques)
// model.scaling(0);
model.primal();
} else {
// barrier
ClpInterior barrier;
barrier.borrowModel(model);
ClpCholeskyBase * cholesky = new ClpCholeskyBase();
cholesky->setKKT(true);
barrier.setCholesky(cholesky);
barrier.primalDual();
barrier.returnModel(model);
}
// Just check if share2qp (quad.mps here)
// this is because I am not checking if variables at ub
if (model.numberColumns() == 79) {
double *primal;
double *dual;
primal = model.primalColumnSolution();
dual = model.dualRowSolution();
// Check duals by hand
const ClpQuadraticObjective * quadraticObj =
(dynamic_cast<const ClpQuadraticObjective*>(model.objectiveAsObject()));
assert(quadraticObj);
CoinPackedMatrix * quad = quadraticObj->quadraticObjective();
const int * columnQuadratic = quad->getIndices();
const CoinBigIndex * columnQuadraticStart = quad->getVectorStarts();
const int * columnQuadraticLength = quad->getVectorLengths();
const double * quadraticElement = quad->getElements();
int numberColumns = model.numberColumns();
int numberRows = model.numberRows();
double * gradient = new double [numberColumns];
// move linear objective
memcpy(gradient, quadraticObj->linearObjective(), numberColumns * sizeof(double));
int iColumn;
for (iColumn = 0; iColumn < numberColumns; iColumn++) {
double valueI = primal[iColumn];
CoinBigIndex j;
for (j = columnQuadraticStart[iColumn];
j < columnQuadraticStart[iColumn] + columnQuadraticLength[iColumn]; j++) {
int jColumn = columnQuadratic[j];
double valueJ = primal[jColumn];
double elementValue = quadraticElement[j];
if (iColumn != jColumn) {
double gradientI = valueJ * elementValue;
double gradientJ = valueI * elementValue;
gradient[iColumn] += gradientI;
gradient[jColumn] += gradientJ;
} else {
double gradientI = valueI * elementValue;
gradient[iColumn] += gradientI;
}
}
if (fabs(primal[iColumn]) > 1.0e-8)
printf("%d primal %g\n", iColumn, primal[iColumn]);
}
for (int i = 0; i < numberRows; i++) {
if (fabs(dual[i]) > 1.0e-8)
printf("%d dual %g\n", i, dual[i]);
}
// Now use duals to get reduced costs
// Can't use this as will try and use scaling
// model.transposeTimes(-1.0,dual,gradient);
// So ...
CoinPackedMatrix * matrix = model.matrix();
const int * row = matrix->getIndices();
const CoinBigIndex * columnStart = matrix->getVectorStarts();
const int * columnLength = matrix->getVectorLengths();
const double * element = matrix->getElements();
for (iColumn = 0; iColumn < numberColumns; iColumn++) {
double dj = gradient[iColumn];
CoinBigIndex j;
for (j = columnStart[iColumn];
j < columnStart[iColumn] + columnLength[iColumn]; j++) {
int jRow = row[j];
dj -= element[j] * dual[jRow];
}
if (model.getColumnStatus(iColumn) == ClpSimplex::basic) {
assert(fabs(dj) < 1.0e-5);
} else {
assert(dj > -1.0e-5);
}
}
delete [] gradient;
}
}
return 0;
}
static int solve(EKKModel * model, int startup, int algorithm,
int presolve)
{
// values pass or not
if (startup)
startup = 1;
// if scaled then be careful
bool scaled = ekk_scaling(model) == 1;
if (scaled)
ekk_scaleRim(model, 1);
void * compressInfo = NULL;
ClpSimplex * clp;
if (!presolve || !presolveInfo) {
// no presolve or osl presolve - compact columns
compressInfo = ekk_compressModel(model);
clp = clpmodel(model, startup);;
} else {
// pick up clp model
clp = presolveInfo->model();
}
// don't scale if alreday scaled
if (scaled)
clp->scaling(false);
if (clp->numberRows() > 10000)
clp->factorization()->maximumPivots(100 + clp->numberRows() / 100);
if (algorithm > 0)
clp->primal(startup);
else
clp->dual();
int numberIterations = clp->numberIterations();
if (presolve && presolveInfo) {
// very wasteful - create a clp copy of osl model
ClpSimplex * clpOriginal = clpmodel(model, 0);
presolveInfo->setOriginalModel(clpOriginal);
// do postsolve
presolveInfo->postsolve(true);
delete clp;
delete presolveInfo;
presolveInfo = NULL;
clp = clpOriginal;
if (presolve == 3 || (presolve == 2 && clp->status())) {
printf("Resolving from postsolved model\n");
clp->primal(1);
numberIterations += clp->numberIterations();
}
}
// put back solution
double * rowDual = (double *) ekk_rowduals(model);
int numberRows = ekk_getInumrows(model);
int numberColumns = ekk_getInumcols(model);
int * rowStatus = (int *) ekk_rowstat(model);
double * rowSolution = (double *) ekk_rowacts(model);
int i;
int * columnStatus = (int *) ekk_colstat(model);
double * columnSolution = (double *) ekk_colsol(model);
memcpy(rowSolution, clp->primalRowSolution(), numberRows * sizeof(double));
memcpy(rowDual, clp->dualRowSolution(), numberRows * sizeof(double));
for (i = 0; i < numberRows; i++) {
if (clp->getRowStatus(i) == ClpSimplex::basic)
rowStatus[i] = 0x80000000;
else
rowStatus[i] = 0;
}
double * columnDual = (double *) ekk_colrcosts(model);
memcpy(columnSolution, clp->primalColumnSolution(),
numberColumns * sizeof(double));
memcpy(columnDual, clp->dualColumnSolution(), numberColumns * sizeof(double));
for (i = 0; i < numberColumns; i++) {
if (clp->getColumnStatus(i) == ClpSimplex::basic)
columnStatus[i] = 0x80000000;
else
columnStatus[i] = 0;
}
ekk_setIprobstat(model, clp->status());
ekk_setRobjvalue(model, clp->objectiveValue());
ekk_setInumpinf(model, clp->numberPrimalInfeasibilities());
ekk_setInumdinf(model, clp->numberDualInfeasibilities());
ekk_setIiternum(model, numberIterations);
ekk_setRsumpinf(model, clp->sumPrimalInfeasibilities());
ekk_setRsumdinf(model, clp->sumDualInfeasibilities());
delete clp;
if (compressInfo)
ekk_decompressModel(model, compressInfo);
if (scaled)
ekk_scaleRim(model, 2);
return 0;
}
int main (int argc, const char *argv[])
{
ClpSimplex model;
int status;
// Keep names
if (argc < 2) {
status = model.readMps("small.mps", true);
} else {
status = model.readMps(argv[1], true);
}
if (status)
exit(10);
/*
This driver implements what I called Sprint. Cplex calls it
"sifting" which is just as silly. When I thought of this trivial idea
it reminded me of an LP code of the 60's called sprint which after
every factorization took a subset of the matrix into memory (all
64K words!) and then iterated very fast on that subset. On the
problems of those days it did not work very well, but it worked very
well on aircrew scheduling problems where there were very large numbers
of columns all with the same flavor.
*/
/* The idea works best if you can get feasible easily. To make it
more general we can add in costed slacks */
int originalNumberColumns = model.numberColumns();
int numberRows = model.numberRows();
// We will need arrays to choose variables. These are too big but ..
double * weight = new double [numberRows+originalNumberColumns];
int * sort = new int [numberRows+originalNumberColumns];
int numberSort = 0;
// Say we are going to add slacks - if you can get a feasible
// solution then do that at the comment - Add in your own coding here
bool addSlacks = true;
if (addSlacks) {
// initial list will just be artificials
// first we will set all variables as close to zero as possible
int iColumn;
const double * columnLower = model.columnLower();
const double * columnUpper = model.columnUpper();
double * columnSolution = model.primalColumnSolution();
for (iColumn = 0; iColumn < originalNumberColumns; iColumn++) {
double value = 0.0;
if (columnLower[iColumn] > 0.0)
value = columnLower[iColumn];
else if (columnUpper[iColumn] < 0.0)
value = columnUpper[iColumn];
columnSolution[iColumn] = value;
}
// now see what that does to row solution
double * rowSolution = model.primalRowSolution();
memset (rowSolution, 0, numberRows * sizeof(double));
model.times(1.0, columnSolution, rowSolution);
int * addStarts = new int [numberRows+1];
int * addRow = new int[numberRows];
double * addElement = new double[numberRows];
const double * lower = model.rowLower();
const double * upper = model.rowUpper();
addStarts[0] = 0;
int numberArtificials = 0;
double * addCost = new double [numberRows];
const double penalty = 1.0e8;
int iRow;
for (iRow = 0; iRow < numberRows; iRow++) {
if (lower[iRow] > rowSolution[iRow]) {
addRow[numberArtificials] = iRow;
addElement[numberArtificials] = 1.0;
addCost[numberArtificials] = penalty;
numberArtificials++;
addStarts[numberArtificials] = numberArtificials;
} else if (upper[iRow] < rowSolution[iRow]) {
addRow[numberArtificials] = iRow;
addElement[numberArtificials] = -1.0;
addCost[numberArtificials] = penalty;
numberArtificials++;
addStarts[numberArtificials] = numberArtificials;
}
}
model.addColumns(numberArtificials, NULL, NULL, addCost,
addStarts, addRow, addElement);
delete [] addStarts;
delete [] addRow;
delete [] addElement;
delete [] addCost;
// Set up initial list
numberSort = numberArtificials;
int i;
for (i = 0; i < numberSort; i++)
sort[i] = i + originalNumberColumns;
} else {
// Get initial list in some magical way
// Add in your own coding here
abort();
}
int numberColumns = model.numberColumns();
const double * columnLower = model.columnLower();
const double * columnUpper = model.columnUpper();
double * fullSolution = model.primalColumnSolution();
// Just do this number of passes
int maxPass = 100;
int iPass;
double lastObjective = 1.0e31;
// Just take this number of columns in small problem
int smallNumberColumns = CoinMin(3 * numberRows, numberColumns);
smallNumberColumns = CoinMax(smallNumberColumns, 3000);
// We will be using all rows
int * whichRows = new int [numberRows];
for (int iRow = 0; iRow < numberRows; iRow++)
whichRows[iRow] = iRow;
double originalOffset;
model.getDblParam(ClpObjOffset, originalOffset);
for (iPass = 0; iPass < maxPass; iPass++) {
printf("Start of pass %d\n", iPass);
//printf("Bug until submodel new version\n");
CoinSort_2(sort, sort + numberSort, weight);
// Create small problem
ClpSimplex small(&model, numberRows, whichRows, numberSort, sort);
// now see what variables left out do to row solution
double * rowSolution = model.primalRowSolution();
memset (rowSolution, 0, numberRows * sizeof(double));
int iRow, iColumn;
// zero out ones in small problem
for (iColumn = 0; iColumn < numberSort; iColumn++) {
int kColumn = sort[iColumn];
fullSolution[kColumn] = 0.0;
}
// Get objective offset
double offset = 0.0;
const double * objective = model.objective();
for (iColumn = 0; iColumn < originalNumberColumns; iColumn++)
offset += fullSolution[iColumn] * objective[iColumn];
small.setDblParam(ClpObjOffset, originalOffset - offset);
model.times(1.0, fullSolution, rowSolution);
double * lower = small.rowLower();
double * upper = small.rowUpper();
for (iRow = 0; iRow < numberRows; iRow++) {
if (lower[iRow] > -1.0e50)
lower[iRow] -= rowSolution[iRow];
if (upper[iRow] < 1.0e50)
upper[iRow] -= rowSolution[iRow];
}
/* For some problems a useful variant is to presolve problem.
In this case you need to adjust smallNumberColumns to get
right size problem. Also you can dispense with creating
small problem and fix variables in large problem and do presolve
on that. */
// Solve
small.primal();
// move solution back
const double * solution = small.primalColumnSolution();
for (iColumn = 0; iColumn < numberSort; iColumn++) {
int kColumn = sort[iColumn];
model.setColumnStatus(kColumn, small.getColumnStatus(iColumn));
fullSolution[kColumn] = solution[iColumn];
}
for (iRow = 0; iRow < numberRows; iRow++)
model.setRowStatus(iRow, small.getRowStatus(iRow));
memcpy(model.primalRowSolution(), small.primalRowSolution(),
numberRows * sizeof(double));
if ((small.objectiveValue() > lastObjective - 1.0e-7 && iPass > 5) ||
!small.numberIterations() ||
iPass == maxPass - 1) {
break; // finished
} else {
lastObjective = small.objectiveValue();
// get reduced cost for large problem
// this assumes minimization
memcpy(weight, model.objective(), numberColumns * sizeof(double));
model.transposeTimes(-1.0, small.dualRowSolution(), weight);
// now massage weight so all basic in plus good djs
for (iColumn = 0; iColumn < numberColumns; iColumn++) {
double dj = weight[iColumn];
double value = fullSolution[iColumn];
if (model.getColumnStatus(iColumn) == ClpSimplex::basic)
dj = -1.0e50;
else if (dj < 0.0 && value < columnUpper[iColumn])
dj = dj;
else if (dj > 0.0 && value > columnLower[iColumn])
dj = -dj;
else if (columnUpper[iColumn] > columnLower[iColumn])
dj = fabs(dj);
else
dj = 1.0e50;
weight[iColumn] = dj;
sort[iColumn] = iColumn;
}
// sort
CoinSort_2(weight, weight + numberColumns, sort);
numberSort = smallNumberColumns;
}
}
if (addSlacks) {
int i;
int numberArtificials = numberColumns - originalNumberColumns;
for (i = 0; i < numberArtificials; i++)
sort[i] = i + originalNumberColumns;
model.deleteColumns(numberArtificials, sort);
}
delete [] weight;
delete [] sort;
delete [] whichRows;
model.primal(1);
return 0;
}