Meta-aprendizagem para seleção de algoritmos sintonizados aplicada ao Problema de Flow Shop Permutacional.

Abstract

Meta-heuristics are high-level search strategies that guide the search for the most promising regions of the solution space and try to escape the optimal local solutions. However, due to the heterogeneity of instances of combinatorial optimization problems, it is not guaranteed that a meta-heuristic can always obtain the best solution in a set of metaheuristics, so algorithm selection, such as meta-learning, can provide a most effective solution when defining which meta-heuristic to choose according to the structural characteristics of the instances. In this work we propose a meta-learning framework for meta-heuristics selection tuned by the race method, which includes meta-features extracted from the graph-based representation of the problem and definition of the meta-classes from of the tuned meta-heuristics. Experiments have shown that the approach is effective for selecting meta-heuristics and their parameters for instances And that the chosen meta-features can extract structural information relevant to the problem addressed.