Originally posted by: SystemAdmin

[shaon said:]

Hi all,

I use the framework of CPLEX 11.1 to develop my own branch and cut algorithm for the mixed binary integer programming problem. I turn off presolve, and use default strategy "MIP emphasis: balance optimality and feasibility."  I have one problem:

In the optimal solution, the values of some binary variables are not 0 or 1, but very close to 0 and 1, e.g., 0.000001 and 0.999997.

Would you please give me some suggestions to figure out this problem?

Thanks.

Shaon

#CPLEXOptimizers
#DecisionOptimization

Originally posted by: SystemAdmin

[prubin said:]

All arithmetic in CPLEX is double-precision, and it's susceptible to rounding error.  I think your best bet is simply to apply a function to the values of binary variables in the solution that rounds them to 0 or 1.

CPLEX has a tolerance parameter (EpInt) that dictates how close to the nearest integer a solution value has to be to be considered integral.  The default value is 1e-5.  You can crank that down if you want.  In fact, CPLEX will accept an EpInt value of 0 (i.e., no rounding error tolerated in integer variables), but I really do not recommend that -- if some rounding error is unavoidable, CPLEX might conclude your problem is infeasible.
#CPLEXOptimizers
#DecisionOptimization