MADS/F-Race: Mesh Adaptive Direct Search Meets F-Race

Finding appropriate parameter settings of parameterized algorithms or AI systems is an ubiquitous task in many practical applications. This task is usually tedious and time-consuming. To reduce human intervention, the study of methods for automated algorithm configuration has received increasing attention in recent years. In this article, we study the mesh adaptive direct search (MADS) method for the configuration of parameterized algorithms. MADS is a direct search method for continuous, global optimization. For handling the stochasticity involved in evaluating the algorithm to be configured, we hybridized MADS with F-Race, a racing method that adaptively allocates an appropriate number of evaluations to each member of a population of candidate algorithm configurations. We experimentally study this hybrid of MADS and F-Race (MADS/F-Race) and compare it to other ways of defining the number of evaluations of each candidate configuration and to another method called I/F-Race. This comparison confirms the good performance and robustness of MADS/F-Race.