This work presents a proposal for automated parameter tuning modeled with a hierarchical optimization structure. A framework is proposed in order to handle the randomness and noise from results for given parameters values of an algorithm of interest. A metaheuristic based on the center of mass concept for solving bilevel optimization called Bilevel Centers Algorithm (BCA) is adapted and tested by configuring representative algorithms for global optimization. Moreover, surrogate models are used to handle the high computational cost inherent to the Automated Parameter Tuning Problem. The approximate model is used to perform a global search to identify promising regions in the parameter search space. The experimental results are interesting and competitive when compared against a state-of-the-art method called irace, which implements a mechanism for reducing computation time.


Release version will be available soon.

Source of the BCAP method to configure algorithms on the branch bcap.