MPILS: An Automatic Tuner for MILP Solvers

The parameter configuration problem consists of finding a parameter configuration that gives a particular algorithm the best performance. This paper introduces a new multi-phase tuner based on the iterated local search meta-heuristic. This tuner addresses the parameter configuration problem for deterministic MILP solvers that are used to solve challenging industrial optimization problems. Further, the proposed tuner offers a new search strategy based on three ideas. First, instead of tuning in the entire configuration space induced by the parameter set, the multi-phase tuner focuses on a small parameter pool that is dynamically enriched with new promising parameters. Second, it leverages the gathered knowledge during the search using statistical learning to forbid less promising parameter combinations. Third, it tunes on a single instance provided by earlier clustering of MILP instances. A computational study on the widely-used commercial solver CPLEX with instances from the MIPLIB library and a real large-scale optimization problem highlights the promising potential of the tuner. • A new MILP tuner to tackle the parameter configuration problem. • It is based on the iterated local search metaheuristic. • It tunes on a small pool of parameters before dynamically adding new promising ones. • It takes profit from the accumulated information using statistical learning. • It yields multiple (quasi-)optimal solutions in few minutes.