Originally posted by: ChristopheV

Dear all,

Yesterday, I've experimented for the first time with the tuning tool. I want to solve a large (see bottom of my message for more details) MIP problem and my time limit is about 4000 seconds. In my case, it seems impossible to find the best solution, but, most of the time, the solutions I get are acceptable for practical purposes.

In the past, I have written my own "tuning tool", which performs runs with large amounts of (combinations of) parameters settings, and returns the best combination of settings for each instance of my problem. In practice, this means that my tool could be running for several days or even weeks. This results in about 10 parameters that should be set to a certain value, resulting in a much better solution/gap than when using the default parameters.

When using the CPLEX tuning tool, I was quite surprised that only a few parameters were tried for my model. In fact, only the following 5 runs were performed:

Test 'defaults':
Test 'easy':
Test 'node_lp_solve':
Test 'aggressive_cuts':
Test 'more_heuristic':

The result of the tuning tool is only the following:

CPLEX Parameter File Version 12.2.0.0
CPX_PARAM_RINSHEUR 100

Is it normal that the tuning tool only performs such a small amount of tests? Is it possible to do more extensive parameter tuning with the tuning tool?

Does anyone have similar of better experiences with the tuning tool? Any suggestions about how to use it?
Thanks!
Christophe
---

Some extra data about my problem:

Reduced MIP has 16512 rows, 155737 columns, and 336798 nonzeros.
Reduced MIP has 3059 binaries, 2301 generals, 0 SOSs, and 0 indicators.
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: SystemAdmin

The tuning tool was designed to produce a reasonable parameter choice with somewhat reasonable effort. It was not designed to use an "infinite" amount of time to return the "optimal" parameter settings.

For this reason, it considers a number of parameter combinations that from our experience are useful for practical problems. Then it tries with quick runs to identify the more promising combinations and finally allows for some longer runs to select the best out of the selected candidates.

Which parameter settings did your tool come up with for your model? Are they really much better than the simple answer returned by the CPLEX tuning tool? And what happens if you compare your parameters with the CPLEX tuning tool parameters on a different model of the same class?

Tobias
#CPLEXOptimizers
#DecisionOptimization