Automated parameter tuning as a bilevel optimization problem solved by a surrogate-assisted population-based approach

摘要

This work presents a proposal for the automated parameter tuning problem (APTP) modeled as a bilevel optimization problem. Different definitions and theoretical results are given in order to formalize the APTP in the context of this hierarchical optimization problem. The obtained bilevel optimization problem is solved via a population-based algorithm added with surrogate models to identify promising regions in the parameter search space. The approach is tested by configuring four representative metaheuristics for numerical optimization on a set of well-known and recent test problems; also a competitive algorithm for a popular combinatorial optimization problem was configured (considering a large benchmark suite). The experimental results are compared against those of a state-of-the-art parameter tuning method called IRACE. The results, validated by the Bayesian signed-rank statistical test, indicate that BCAP, even it is based on an usually costly model (i.e. a bilevel optimization problem), with only half of the calls to the target algorithm, is able to find better configurations than those obtained by the compared approach.