Originally posted by: ORA17

Step 1:

I am solving a mixed integer programming problem with N variables and M constraints. And I want to select K best integer variable that gives maximum objective value.

I am using param: cplex.SetParam(Cplex.IntParam.RootAlg, 4)

Step 2:

I am solving same problem (i.e. with same set of constraints) but with K variables (K <= N) which is subset of N variables and output of Step 1 but getting different (mostly higher) objective value.

I am using param: cplex.SetParam(Cplex.IntParam.RootAlg, 4)

But If I remove cplex.SetParam(Cplex.IntParam.RootAlg, 4) from both  I get same objective value in both of my problems.

please note that this issue I face rarely not always.

Could you please help me to check why I am facing this kind of issues?

What is the difference in the objective value between two offending runs? Is it between the relative MIP gap tolerance? Does the problem persist if you set IloCplex.NumParam.EpGap to 0?

Can you attach a pair of offending models? Or at least show the logs for the two solves?

Originally posted by: ORA17

Hi @DanielJunglas 548c249b-bb6b-42f6-99b8-f0376c80829e  Thank you for your kind reply.

I have checked my cplex log and observed that when I am using barrier to solve my MIP, cplex gives me fractional solution. And with default settings cplex gives integer solution.

Problem is in my programming logic I tried to convert fractional solution to integer solution without rounding it. But both giving same objective function.

But still I have question why am I getting fraction solution when I solve MIP?

I have attached my cplex log for your kind review. Thanks.

How much deviate the fractional values from integer values? Small round-off is expected and cannot be avoided unless you use a solver that implements arbitrary precision arithmetic. Usually slightly fractional values are not a problem and rounding them is fine. You can set IloCplex.NumParam.EpInt to 0 to make it less likely for these values to appear. This may however hurt performance in some cases.