Performance MIPstart

Specific Domains Optimization (Mathematical)
1

I’ll share an experience.
I have a mixed linear optimization problem.
When I use a feasibre solution (MIPstart), sometimes CPLEX performs worse than if I let it determine an incumbent solution automatically.
Sometimes when I don’t use MIPstart, CPLEX doesn’t even determine an integer solution.

for i in I
    for t in T
        set_start_value(s[i,t],1)
    end
end

What is your experience in this matter? Fix variables and ensure an incumbet solution or leave CPLEX determine automatically?

2

When you supply a feasible solution, is it complete? That is, do you provide start values for all variables, or only a subset?

If you only provide a partial solution, than many solvers will try to solve the subproblem where those are fixed. This might take significant time and not be worth it if your candidate is of poor quality.

Otherwise, what you observe might be covered by the phenomen of “performance variability”, typical in MIP solvers.

1 Like
3

I provide only the binary variables of my problem. A good advise is supply all variables. I will test.