Subject: CPLEX vs Gurobi: Feasible but Better Solution from CPLEX?

GAMS
, ,
1

Hi everyone,
I’ve been solving a maximization model using both CPLEX and Gurobi and got an unexpected result.

I set a time limit of 5 minutes for CPLEX, and it returned:

SOLVER STATUS: 3 (Resource Interrupt) MODEL STATUS: 8 (Integer Solution)

OBJECTIVE VALUE: 25,123,015,162.5282

Then I solved the same model using Gurobi, which finished in just 3 seconds:

SOLVER STATUS: 1 (Normal Completion) MODEL STATUS: 1 (Optimal)

OBJECTIVE VALUE: 22,911,669,452.9621

What puzzles me is that CPLEX feasible solution has a significantly higher objective value than Gurobi optimal one. I’m wondering:
Could CPLEX have found a better solution that Gurobi missed?
Or is it more likely that CPLEX solution is infeasible or incorrect in some way, even though it’s reported as an “Integer Solution”?
Has anyone seen this kind of discrepancy before?
Which solution would you consider more reliable in this situation?
Any help or insight would be greatly appreciated.
Thanks!

2

If solvers behave inconsistently, the model is often numerically challenging (e.g. because it is poorly scaled).

Is it possible to share your model, such that other users can reproduce the issue and maybe come up with suggestions?