Fine-Tuning Algorithm Parameters Using the Design of Experiments Approach

Abstract

Optimizing parameter settings is an important task in algorithm design. Several automated parameter tuning procedures/configurators have been proposed in the literature, most of which work effectively when given a good initial range for the parameter values. In the Design of Experiments (DOE), a good initial range is known to lead to an optimum parameter setting. In this paper, we present a framework based on DOE to find a good initial range of parameter values for automated tuning. We use a factorial experiment design to first screen and rank all the parameters thereby allowing us to then focus on the parameter search space of the important parameters. A model based on the Response Surface methodology is then proposed to define the promising initial range for the important parameter values. We show how our approach can be embedded with existing automated parameter tuning configurators, namely ParamILS and RCS (Randomized Convex Search), to tune target algorithms and demonstrate that our proposed methodology leads to improvements in terms of the quality of the solutions.