Originally posted by: theog

Hello,

I have implemented a linear MIP model using the CPLEX concert technology in C++ and Java.

I have also analyzed output results after running both models in some instances (e.g., number of nodes solved by branch-and-bound).

The results differ from each other. More specifically, the number of nodes solved by MIP-CPLEX-Java is much less than those of  MIP-CPLEX-C++.

The question is: Is it correct? Does the implementation of each library differs from each other (presolve optimizations or optimizations performed in execution time), or maybe I made mistakes in the C++ code?

Best regards,

theog

Before calling solve, in each program, export your problem to SAV format.  Try running each of these in the CPLEX interactive.  Do you get the same results for each there?

Originally posted by: IrvL

CPLEX converts the object-oriented representations in C++ Concert and Java Concert differently, which can cause the variables and constraints to be in a different order. That ordering can cause different performance in the underlying solver.

Having said that, compare the log files and see if the two implementations are truly identical.  The number of constraints, variables, nonzeros should be the same. The minimum/maximum values of the matrix and bounds and objective coefficients should be the same.  If not, then you have a bug somewhere....

-Irv Lustig

Optimization Principal

Princeton Consultants