Cbc

Getting Good Bounds in CBC

CbcHeuristic - Heuristic Methods

In practice, it is very useful to get a good solution reasonably fast. Any MIP-feasible solution produces an upper bound, and a good bound will greatly reduce the run time. Good solutions can satisfy the user on very large problems where a complete search is impossible. Obviously, heuristics are problem dependent, although some do have more general use. At present there is only one heuristic in CBC itself, CbcRounding. Hopefully, the number will grow. Other heuristics are in the Cbc/examples directory. A heuristic tries to obtain a solution to the original problem so it only needs to consider the original rows and does not have to use the current bounds. CBC provides an abstract base class CbcHeuristic and a rounding heuristic in CBC.

This chapter describes how to build a greedy heuristic for a set covering problem, e.g., the miplib problem fast0507. A more general (and efficient) version of the heuristic is in CbcHeuristicGreedy.hpp and CbcHeuristicGreedy.cpp located in the Cbc/examples directory.

The greedy heuristic will leave all variables taking value one at this node of the tree at value one, and will initially set all other variable to value zero. All variables are then sorted in order of their cost divided by the number of entries in rows which are not yet covered. (We may randomize that value a bit so that ties will be broken in different ways on different runs of the heuristic.) The best one is choosen, and set to one. The process is repeated. Because this is a set covering problem (i.e., all constraints are GE), the heuristic is guaranteed to find a solution (but not necessarily an improved solution). The speed of the heuristic could be improved by just redoing those affected, but for illustrative purposes we will keep it simple.(The speed could also be improved if all elements are 1.0).

The key CbcHeuristic method is int solution(double & solutionValue, double * betterSolution). The solution() method returns 0 if no solution found, and returns 1 if a solution is found, in which case it fills in the objective value and primal solution. The code in CbcHeuristicGreedy.cpp is a little more complicated than this following example. For instance, the code here assumes all variables are integer. The important bit of data is a copy of the matrix (stored by column) before any cuts have been made. The data used are bounds, objective and the matrix plus two work arrays.

OsiSolverInterface * solver = model_->solver(); // Get solver from CbcModel
const double * columnLower = solver->getColLower(); // Column Bounds
const double * columnUpper = solver->getColUpper();
const double * rowLower = solver->getRowLower(); // We know we only need lower bounds
const double * solution = solver->getColSolution();
const double * objective = solver->getObjCoefficients(); // In code we also use min/max
double integerTolerance = model_->getDblParam(CbcModel::CbcIntegerTolerance);
double primalTolerance;
solver->getDblParam(OsiPrimalTolerance,primalTolerance);
int numberRows = originalNumberRows_; // This is number of rows when matrix was passed in
// Column copy of matrix (before cuts)
const double * element = matrix_.getElements();
const int * row = matrix_.getIndices();
const CoinBigIndex * columnStart = matrix_.getVectorStarts();
const int * columnLength = matrix_.getVectorLengths();

// Get solution array for heuristic solution
int numberColumns = solver->getNumCols();
double * newSolution = new double [numberColumns];
// And to sum row activities
double * rowActivity = new double[numberRows];

The newSolution is then initialized to the rounded down solution:

for (iColumn=0;iColumn<numberColumns;iColumn++) {
  CoinBigIndex j;
  double value = solution[iColumn];
  // Round down integer
  if (fabs(floor(value+0.5)-value)<integerTolerance)
    value=floor(CoinMax(value+1.0e-3,columnLower[iColumn]));
  // make sure clean
  value = CoinMin(value,columnUpper[iColumn]);
  value = CoinMax(value,columnLower[iColumn]);
  newSolution[iColumn]=value;
  if (value) {
    double cost = direction * objective[iColumn];
    newSolutionValue += value*cost;
    for (j=columnStart[iColumn];j<columnStart[iColumn]+columnLength[iColumn];j++) {
      int iRow=row[j];
      rowActivity[iRow] += value*element[j];
    }
  }
}

At this point some row activities may be below their lower bound. To correct this infeasibility, the variable which is cheapest in reducing the sum of infeasibilities is found and updated, and the process repeats. This is a finite process. (Theimplementation could be faster, but is kept simple for illustrative purposes.)

while (true) {
  // Get column with best ratio
  int bestColumn=-1;
  double bestRatio=COIN_DBL_MAX;
  for (int iColumn=0;iColumn<numberColumns;iColumn++) {
    CoinBigIndex j;
    double value = newSolution[iColumn];
    double cost = direction * objective[iColumn];
    // we could use original upper rather than current
    if (value+0.99<columnUpper[iColumn]) {
      double sum=0.0; // Compute how much we will reduce infeasibility by
      for (j=columnStart[iColumn];
           j<columnStart[iColumn]+columnLength[iColumn];j++) {
        int iRow=row[j];
        double gap = rowLower[iRow]-rowActivity[iRow];
        if (gap>1.0e-7)
          sum += CoinMin(element[j],gap);
        if (element[j]+rowActivity[iRow]<rowLower[iRow]+1.0e-7)
          sum += element[j];
      }
      if (sum>0.0) {
        double ratio = (cost/sum)*(1.0+0.1*CoinDrand48());
        if (ratio<bestRatio) {
          bestRatio=ratio;
          bestColumn=iColumn;
        }
      }
    }
  }
  if (bestColumn<0)
    break; // we have finished
  // Increase chosen column
  newSolution[bestColumn] += 1.0;
  double cost = direction * objective[bestColumn];
  newSolutionValue += cost;
  for (CoinBigIndex j=columnStart[bestColumn];
    j<columnStart[bestColumn]+columnLength[bestColumn];j++) {
    int iRow = row[j];
    rowActivity[iRow] += element[j];
  }
}

A solution value of newSolution is compared to the best solution value. If newSolution is an improvement, its feasibility is validated.

returnCode=0; // 0 means no good solution
if (newSolutionValue<solutionValue) { // minimization
  // check feasible
  memset(rowActivity,0,numberRows*sizeof(double));
  for (iColumn=0;iColumn<numberColumns;iColumn++) {
    CoinBigIndex j;
    double value = newSolution[iColumn];
    if (value) {
      for (j=columnStart[iColumn];
           j<columnStart[iColumn]+columnLength[iColumn];j++) {
        int iRow=row[j];
        rowActivity[iRow] += value*element[j];
      }
    }
  }
  // check was approximately feasible
  bool feasible=true;
  for (iRow=0;iRow<numberRows;iRow++) {
    if(rowActivity[iRow]<rowLower[iRow])
      if (rowActivity[iRow]<rowLower[iRow]-10.0*primalTolerance)
        feasible = false;
  }
  if (feasible) {
    // new solution
    memcpy(betterSolution,newSolution,numberColumns*sizeof(double));
    solutionValue = newSolutionValue;
    // We have good solution
    returnCode=1;
  }
}

Notes on the Feasibility Pump Implementation

References are

The basic idea (much simplified) is that you start with the relaxed continuous solution and then keep changing the objective function to try and minimize the sum of integer infeasibilities. If this goes to zero then you have a solution.

So, how long do you try? If you get a solution, can you get a better one? Even if you fail to get a solution, can you get any benefit? The answer to the second question is yes - by adding a constraint that says the objective must be better than this solution. The answer to the third question is also yes. By continually changing the objective, you are making the solution go all over the place but there will be variables which despite all this remain fixed at a value. Maybe if you fix these variables and preprocess the resulting problem you will get a much smaller problem on which it will be worth doing a few nodes of branch and cut.

By default the Feasibility Pump heuristic is run at the start of CBC’s branch and cut, but there are several ways to fine tune the heuristic. Before I answer any of those questions, I should mention one or two undocumented features. There are several parameters you can set to terminate the search early. One is maxNodes. If you do maxN?? in Cbc it will say that the valid range is -1 to 2147483647. If you set maxNodes to 0, cbc will apply all cuts at root node and then terminate, but if you set it to -1, then the cuts will not be computed. This can save some time if all you want to do is find a heuristic solution. Also if allowableGap or ratioGap is set and a heuristic finds a solution which satisfies the gap criterion, then the code will also terminate before doing the node 0 cut computations. Finally I should mention the doh option. The full name of the parameter is doHeuristic. Normally Cbc does preprocessing and then enters branch and cut, where the first thing to be done is the heuristics phase. If a valid value for the objective is known then sometimes the preprocessing can do a better job as it can fix variables on reduced costs. One way of doing this is to run some heuristics, then do preprocessing and go into branch and cut. To do the feasibility pump then do preprocessing and then do feasibility pump inside branch and cut one would do

cbc ....mps -feas both -doh -solve

Anyway back to tuning the feasibility pump. I should mention here that I will give the full name of the option but you can abbreviate it e.g. passFeasibilityPump is accepted if you enter passF. There are three options which affect the heuristic:

This site is open source. Improve this page.