Originally posted by: gabb

Hi,

I have built an MIP model and coded the problem using Java concert technology. Because the size of the problem is really huge, I set the running time by the following sentence.

cplex.setParam(DoubleParam.TiLim, CPLEX_TIME_LIMIT);

and set CPLEX_TIME_LIMIT to 60 seconds.

I ran the same model with same input data twice but got different solutions and log info. The "best integer", "best bound", and "gap" are different in two runs. Below is part of the log info.

Run #1:

Run #2:

Why are they different? Is it possible to get identical solutions at different runs when solving a large-scale problem to a near-optimal solution (terminate before the solver finding the real optimal solution)?

Thank you very much!

There are at least two sources of variability (maybe more). The first is that CPLEX defaults to parallel threading. If you run the same job twice but the load on the system (from everything other than CPLEX) changes -- as it most definitely will -- your log may change a bit.This is very likely also true when you force CPLEX to run single-threaded, but my feeling is that multiple threads increase variability.

The second, and likely greater, source of variability is that CPLEX internally makes some decisions "randomly" (using a pseudorandom number generator). You can account for this by setting the seed for the random number generator. The Java name for the (integer) seed parameter is IloCplex.Param.RandomSeed.

Originally posted by: gabb

Thank you very much for your reply!

I will run some tests on single thread and fix the seed parameter to see if things work out.

Another source for non-deterministic behavior is that time limit. Try to run with a deterministic time limit or with another deterministic limit such as node count or gap.

Originally posted by: gabb

Thank you!