// Creates a branching object from this infeasible object.
BcpsBranchObject *
BlisObjectInt::createBranchObject(BcpsModel *m, int direction) const
{
BlisModel *model = dynamic_cast<BlisModel* >(m);
OsiSolverInterface * solver = model->solver();
double integerTolerance = model->BlisPar()->entry(BlisParams::integerTol);
const double * solution = solver->getColSolution();
const double * lower = solver->getColLower();
const double * upper = solver->getColUpper();
double value = solution[columnIndex_];
//std::cout << "COL"<< columnIndex_ << ": x = " << value << std::endl;
// Force value in bounds.
value = CoinMax(value, lower[columnIndex_]);
value = CoinMin(value, upper[columnIndex_]);
double nearest = floor(value + 0.5);
assert (upper[columnIndex_] > lower[columnIndex_]);
int hotstartStrategy = model->getHotstartStrategy();
if (hotstartStrategy <= 0) {
if (fabs(value - nearest) < integerTolerance) {
// Already integeral.
std::cout << "ERROR: COL" << columnIndex_ << ": x=" << value
<< ", nearest=" << nearest
<< ", intTol=" << integerTolerance << std::endl;
assert(0);
}
}
else {
const double * incumbent = model->incumbent();
double targetValue = incumbent[columnIndex_];
if (direction > 0) {
value = targetValue - 0.1;
}
else {
value = targetValue + 0.1;
}
}
return new BlisBranchObjectInt(model, objectIndex_, direction, value);
}
// Compute the infeasibility based on currently relax solution.
double
BlisObjectInt::infeasibility(BcpsModel *m, int & preferredWay) const
{
BlisModel * model = dynamic_cast<BlisModel *>(m);
OsiSolverInterface * solver = model->solver();
const double * solution = solver->getColSolution();
const double * lower = solver->getColLower();
const double * upper = solver->getColUpper();
double value = solution[columnIndex_];
value = std::max(value, lower[columnIndex_]);
value = std::min(value, upper[columnIndex_]);
//printf("%d %g %g %g %g\n",columnIndex_,value,lower[columnIndex_],
// solution[columnIndex_],upper[columnIndex_]);
double nearest = floor(value + (1.0 - breakEven_));
double integerTolerance = model->BlisPar()->entry(BlisParams::integerTol);
if (nearest > value) {
preferredWay = 1;
}
else {
preferredWay = -1;
}
double weight = fabs(value - nearest);
// normalize so weight is 0.5 at break even
if (nearest < value) {
weight = (0.5/breakEven_) * weight;
}
else {
weight = (0.5/(1.0 - breakEven_)) * weight;
}
if (fabs(value - nearest) <= integerTolerance) {
return 0.0;
}
else {
return weight;
}
}
BcpsBranchObject *
BlisObjectInt::preferredNewFeasible(BcpsModel *m) const
{
BlisModel *model = dynamic_cast<BlisModel* >(m);
OsiSolverInterface * solver = model->solver();
double value = solver->getColSolution()[columnIndex_];
double nearest = floor(value + 0.5);
double integerTolerance = model->BlisPar()->entry(BlisParams::integerTol);
assert(fabs(value - nearest) <= integerTolerance);
double dj = solver->getObjSense()*solver->getReducedCost()[columnIndex_];
BlisBranchObjectInt * object = NULL;
if (dj >= 0.0) {
// Better go down
if (nearest > originalLower_ + 0.5) {
// Has room to go down
object = new BlisBranchObjectInt(model,
objectIndex_,
-1,
nearest - 1.0,
nearest - 1.0);
}
}
else {
// Better go up
if (nearest < originalUpper_ - 0.5) {
// Has room to go up
object = new BlisBranchObjectInt(model,
objectIndex_,
-1,
nearest + 1.0,
nearest + 1.0);
}
}
return object;
}
/** Create a set of candidate branching objects. */
int
BlisBranchStrategyRel::createCandBranchObjects(int numPassesLeft)
{
int bStatus = 0;
int i, pass, colInd;
int preferDir, saveLimit;
int numFirsts = 0;
int numInfs = 0;
int minCount = 0;
int numLowerTightens = 0;
int numUpperTightens = 0;
double lpX, score, infeasibility, downDeg, upDeg, sumDeg = 0.0;
bool roundAgain, downKeep, downGood, upKeep, upGood;
int *lbInd = NULL;
int *ubInd = NULL;
double *newLB = NULL;
double *newUB = NULL;
double * saveUpper = NULL;
double * saveLower = NULL;
double * saveSolution = NULL;
BlisModel *model = dynamic_cast<BlisModel *>(model_);
OsiSolverInterface * solver = model->solver();
int numCols = model->getNumCols();
int numObjects = model->numObjects();
//int lookAhead = dynamic_cast<BlisParams*>
// (model->blisPar())->entry(BlisParams::lookAhead);
//------------------------------------------------------
// Check if max time is reached or no pass is left.
//------------------------------------------------------
double timeLimit = model->AlpsPar()->entry(AlpsParams::timeLimit);
bool maxTimeReached = (CoinCpuTime() - model->startTime_ > timeLimit);
bool selectNow = false;
if (maxTimeReached || !numPassesLeft) {
selectNow = true;
#ifdef BLIS_DEBUG
printf("REL: CREATE: maxTimeReached %d, numPassesLeft %d\n",
maxTimeReached, numPassesLeft);
#endif
}
// Store first time objects.
std::vector<BlisObjectInt *> firstObjects;
// Store infeasible objects.
std::vector<BlisObjectInt *> infObjects;
// TODO: check if sorting is expensive.
std::multimap<double, BlisObjectInt*, BlisPseuoGreater> sortedObjects;
double objValue = solver->getObjSense() * solver->getObjValue();
const double * lower = solver->getColLower();
const double * upper = solver->getColUpper();
int lookAhead = dynamic_cast<BlisParams*>
(model->BlisPar())->entry(BlisParams::lookAhead);
BlisObjectInt * intObject = NULL;
//------------------------------------------------------
// Backup solver status and mark hot start.
//-----------------------------------------------------
saveSolution = new double[numCols];
memcpy(saveSolution, solver->getColSolution(), numCols*sizeof(double));
saveLower = new double[numCols];
saveUpper = new double[numCols];
memcpy(saveLower, lower, numCols * sizeof(double));
memcpy(saveUpper, upper, numCols * sizeof(double));
//------------------------------------------------------
// Find the infeasible objects.
// NOTE: we might go round this loop twice if we are feed in
// a "feasible" solution.
//------------------------------------------------------
for (pass = 0; pass < 2; ++pass) {
numInfs = 0;
BcpsObject * object = NULL;
infObjects.clear();
firstObjects.clear();
for (i = 0; i < numObjects; ++i) {
object = model->objects(i);
infeasibility = object->infeasibility(model, preferDir);
if (infeasibility) {
++numInfs;
intObject = dynamic_cast<BlisObjectInt *>(object);
if (intObject) {
//score = object->pseudocost().getScore();
//tempBO = object->createBranchObject(model, preferDir);
//candObjects.insert(std::make_pair(score, tempBO));
//tempBO = NULL;
infObjects.push_back(intObject);
if (!selectNow) {
minCount =
ALPS_MIN(intObject->pseudocost().getDownCount(),
intObject->pseudocost().getUpCount());
if (minCount < 1) {
firstObjects.push_back(intObject);
}
}
#ifdef BLIS_DEBUG_MORE
if (intObject->columnIndex() == 15) {
std::cout << "x[15] = " << saveSolution[15]
<< std::endl;
}
#endif
intObject = NULL;
}
else {
// TODO: currently all are integer objects.
#ifdef BLIS_DEBU
assert(0);
#endif
}
}
}
if (numInfs) {
#ifdef BLIS_DEBUG_MORE
std::cout << "REL: numInfs = " << numInfs
<< std::endl;
#endif
break;
}
else if (pass == 0) {
// The first pass and is IP feasible.
#ifdef BLIS_DEBUG
std::cout << "REL: given a feasible sol" << std::endl;
#endif
roundAgain = false;
CoinWarmStartBasis * ws =
dynamic_cast<CoinWarmStartBasis*>(solver->getWarmStart());
if (!ws) break;
// Force solution values within bounds
for (i = 0; i < numCols; ++i) {
lpX = saveSolution[i];
if (lpX < lower[i]) {
saveSolution[i] = lower[i];
roundAgain = true;
ws->setStructStatus(i, CoinWarmStartBasis::atLowerBound);
}
else if (lpX > upper[i]) {
saveSolution[i] = upper[i];
roundAgain = true;
ws->setStructStatus(i, CoinWarmStartBasis::atUpperBound);
}
}
if (roundAgain) {
// Need resolve and do the second round selection.
solver->setWarmStart(ws);
delete ws;
// Resolve.
solver->resolve();
if (!solver->isProvenOptimal()) {
// Become infeasible, can do nothing.
bStatus = -2;
goto TERM_CREATE;
}
else {
// Save new lp solution.
memcpy(saveSolution, solver->getColSolution(),
numCols * sizeof(double));
objValue = solver->getObjSense() * solver->getObjValue();
}
}
else {
delete ws;
break;
}
}
} // EOF 2 pass
//--------------------------------------------------
// If we have a set of first time object,
// branch up and down to initialize pseudo-cost.
//--------------------------------------------------
numFirsts = static_cast<int> (firstObjects.size());
if (numFirsts > 0) {
CoinWarmStart * ws = solver->getWarmStart();
solver->getIntParam(OsiMaxNumIterationHotStart, saveLimit);
int maxIter = ALPS_MAX(model->getAveIterations(), 50);
solver->setIntParam(OsiMaxNumIterationHotStart, maxIter);
solver->markHotStart();
lbInd = new int [numFirsts];
ubInd = new int [numFirsts];
newLB = new double [numFirsts];
newUB = new double [numFirsts];
for (i = 0; i < numFirsts && bStatus != -2; ++i) {
colInd = firstObjects[i]->columnIndex();
lpX = saveSolution[colInd];
BlisStrongBranch(model, objValue, colInd, lpX,
saveLower, saveUpper,
downKeep, downGood, downDeg,
upKeep, upGood, upDeg);
if(!downKeep && !upKeep) {
// Both branch can be fathomed
bStatus = -2;
}
else if (!downKeep) {
// Down branch can be fathomed.
lbInd[numLowerTightens] = colInd;
newLB[numLowerTightens++] = ceil(lpX);
//break;
}
else if (!upKeep) {
// Up branch can be fathomed.
ubInd[numUpperTightens] = colInd;
newUB[numUpperTightens++] = floor(lpX);
// break;
}
// Update pseudocost.
if(downGood) {
firstObjects[i]->pseudocost().update(-1, downDeg, lpX);
}
if(downGood) {
firstObjects[i]->pseudocost().update(1, upDeg, lpX);
}
}
//--------------------------------------------------
// Set new bounds in lp solver for resolving
//--------------------------------------------------
if (bStatus != -2) {
if (numUpperTightens > 0) {
bStatus = -1;
for (i = 0; i < numUpperTightens; ++i) {
solver->setColUpper(ubInd[i], newUB[i]);
}
}
if (numLowerTightens > 0) {
bStatus = -1;
for (i = 0; i < numLowerTightens; ++i) {
solver->setColLower(lbInd[i], newLB[i]);
}
}
}
//--------------------------------------------------
// Unmark hotstart and recover LP solver.
//--------------------------------------------------
solver->unmarkHotStart();
solver->setColSolution(saveSolution);
solver->setIntParam(OsiMaxNumIterationHotStart, saveLimit);
solver->setWarmStart(ws);
delete ws;
}
//std::cout << "REL: bStatus = " << bStatus << std::endl;
if (bStatus < 0) {
// Infeasible or monotone.
goto TERM_CREATE;
}
else {
// All object's pseudocost have been initialized.
// Sort them, and do strong branch for the unreliable one
// NOTE: it set model->savedLpSolution.
// model->feasibleSolution(numIntegerInfs, numObjectInfs);
sumDeg = 0.0;
for (i = 0; i < numInfs; ++i) {
score = infObjects[i]->pseudocost().getScore();
sumDeg += score;
std::pair<const double, BlisObjectInt*> sa(score, infObjects[i]);
sortedObjects.insert(sa);
#ifdef BLIS_DEBUG_MORE
std::cout << "col[" << infObjects[i]->columnIndex() << "]="
<< score << ", "<< std::endl;
#endif
}
int numNotChange = 0;
std::multimap< double, BlisObjectInt*, BlisPseuoGreater >::iterator pos;
CoinWarmStart * ws = solver->getWarmStart();
solver->getIntParam(OsiMaxNumIterationHotStart, saveLimit);
int maxIter = ALPS_MAX(model->getAveIterations(), 50);
solver->setIntParam(OsiMaxNumIterationHotStart, maxIter);
solver->markHotStart();
BlisObjectInt *bestObject = NULL;
double bestScore = -10.0;
for (pos = sortedObjects.begin(); pos != sortedObjects.end(); ++pos) {
intObject = pos->second;
colInd = intObject->columnIndex();
#ifdef BLIS_DEBUG_MORE
std::cout << "col[" << colInd << "]: "
<< "score=" << pos->first
<< ", upCount=" << intObject->pseudocost().getUpCount()
<<", downCount="<< intObject->pseudocost().getDownCount()
<< std::endl;
#endif
// Check if reliable.
int objRelibility=ALPS_MIN(intObject->pseudocost().getUpCount(),
intObject->pseudocost().getDownCount());
if (objRelibility < relibility_) {
// Unrelible object. Do strong branching.
lpX = saveSolution[colInd];
BlisStrongBranch(model, objValue, colInd, lpX,
saveLower, saveUpper,
downKeep, downGood, downDeg,
upKeep, upGood, upDeg);
// Update pseudocost.
if(downGood) {
intObject->pseudocost().update(-1, downDeg, lpX);
}
if(downGood) {
intObject->pseudocost().update(1, upDeg, lpX);
}
}
// Compare with the best.
if (intObject->pseudocost().getScore() > bestScore) {
bestScore = intObject->pseudocost().getScore();
bestObject = intObject;
// Reset
numNotChange = 0;
}
else {
// If best doesn't change for "lookAhead" comparisons, then
// the best is reliable.
if (++numNotChange > lookAhead) {
if (bestObject->pseudocost().getUpCost() >
bestObject->pseudocost().getDownCost()) {
preferDir = 1;
}
else {
preferDir = -1;
}
break;
}
}
}
solver->unmarkHotStart();
solver->setColSolution(saveSolution);
solver->setIntParam(OsiMaxNumIterationHotStart, saveLimit);
solver->setWarmStart(ws);
delete ws;
model->setSolEstimate(objValue + sumDeg);
assert(bestObject != NULL);
bestBranchObject_ = bestObject->createBranchObject(model, preferDir);
}
TERM_CREATE:
//------------------------------------------------------
// Cleanup.
//------------------------------------------------------
delete [] lbInd;
delete [] ubInd;
delete [] newLB;
delete [] newUB;
delete [] saveSolution;
delete [] saveLower;
delete [] saveUpper;
return bStatus;
}