Abstract

Metaheuristics have been successfully applied to solve complex real-world problems in many application domains. Their performance strongly depends on the values of their parameters. Many tuning algorithms have already been proposed to find a set of suitable values. However, the amount of computational time required to obtain these values is usually high. Our objective is to propose a collaborative strategy to: (1) improve the quality of configurations obtained by tuner algorithms and (2) reduce the time consumed in the tuning process. Here, we introduce a novel opposite scoring (OS) strategy that learns from configurations that produce a positive and a negative effect in the target algorithm. However, OS guides its trajectory by choosing parameter configurations that decrease the performance of the target algorithm. For the learning process, OS stores the quality of all the evaluated configurations and computes a score for each value in the visited parameter configurations. Then, OS generates the initial set of configurations for a tuner, where values that obtain a better score will have a higher probability of being part of this set. We evaluate our proposal using the well-known Evolutionary Calibrator (Evoca). Also, we tune three different algorithms: an Ant Colony Optimization algorithm for solving the Multidimensional Knapsack Problem, a Genetic Algorithm for solving landscapes that follow the NK model (N components and degree K), and a Particle Swarm Optimization algorithm for solving continuous optimization problems. Results show that OS-Evoca obtains better quality configurations than Evoca, consuming less computational resources.