Originally posted by: x3t2h

Hi,

in order to compare the performance of my own Branch and Bound formulation with the built in MIP solver, I am searching for relevant paramters to make the MIP solver use a "pure" branch and bound algorithm, i.e no heuristics, cuts etc. I am using "depth first" node selection and "most infeasible" variable selection in my formulation. The settings I am using so far are:

• Use dual simplex: setParam(IloCplex::RootAlg, 2)
•  Deactivate presolver: setParam(IloCplex::PreInd, 0);
• Deactive cuts: setParam(IloCplex::NumParam::CutsFactor, 1)
• Branch on most infeasible variable: setParam(IloCplex::Param::MIP::Strategy::VariableSelect, 1)
• Depth first node selection: setParam(IloCplex::Param::MIP::Strategy::NodeSelect, 0)
• Disable dynamic search: setParam(IloCplex::Param::MIP::Strategy::Search, 1)
• Limit threads to 1: setParam(IloCplex::Param::Threads, 1)
• Reduce backtracking: setParam(IloCplex::Param::MIP::Strategy::Backtrack, 1) (It is not completely turned off, right?)
• Turn off periodic heuritic: setParam(IloCplex::Param::MIP::Strategy::HeuristicFreq, -1)
• Do not apply RINS heuristic: setParam(IloCplex::Param::MIP::Strategy::RINSHeur, -1)
• Do not repeat presolve: setParam(IloCplex::Param::Preprocessing::RepeatPresolve, 0)

Is there anything else I could use in order to turn off as many "advanced features" of Cplex as possible and make the MIP solver use the most basic branch and bound version?         

Thanks a lot for your help!

This seems a good start. For cuts it is however to use parameter CutPass ot make sure CPLEX does not perform any cutting plane passes.

However, keep this in mind:

• CPLEX is designed to be fast and smart. So there are things that you just cannot disable.
• A good number of things that make CPLEX fast are trade secrets, hence there are no parameters anyway.

Consequently, it is not possible to have CPLEX perform B&B in the "dumbest form possible". It will always do something more.

Also, why limit threads to 1? That is not necessary to get as close as possible to "pure" B&B.

Originally posted by: x3t2h

Thanks a lot. I limited the threads to 1 as my own branch and bound implementation is currently only capable of using one thread, but of course it is not necessary to do that just to obtain a "pure" branch and bound algorithm.