Abstract

Most of the algorithms designed for problem solving have many parameters which values determine their performance. Tuning methods or calibrators are algorithms whose goal is to automate the process of selecting the parameter values of heuristic based algorithms to efficiently solve complex search problems. However, many algorithms are still tuned by-hand either because of the execution time required or the number of scenarios to define before a calibrator is executed. In this work, we propose a practical tuning method that uses a local search procedure that allows obtaining good calibrations in a reduced amount of time, compared to other well-known calibrators. Our tuner has an opposite-inspired learning component used to focus on the most promising areas of the parameter values search space and gathers useful parameter information that is provided to the user. We compare our proposal with two well-known tuners to calibrate two classical optimization problems. We also evaluate the relevance of the opposite-inspired learning component during the search process. A convergence and statistical analysis are presented to confirm that our approach is a good option especially when the user does not have enough time for tuning.