Algoritmo evolutivo multi-objetivo baseado em decomposição com arquivo externo e adaptação de pesos baseada em vizinhança local

Abstract

Multiobjective evolutionary algorithms (MOEA) present an interesting approach to solving different types of problems, known as multiobjective problems (MOP). The subcategory of MOEA with decomposition-based methods have been growing rapidly and many studies have shown that the distribution of weight vectors plays an interesting factor to obtain a uniform set of solutions. However, an uniform distribution of weight vectors at the beginning of evolution not always result in an uniform set of solutions in the objective space, as the results are highly dependent on the Pareto front shape. Irregularly shaped Pareto fronts (disconnected, inverted, etc.) generaly do not contains all parts of the initial set of weight vectors. One approach to overcome this problem is to adapt the weight vectors to approximate the shape of the Pareto boundary. Aiming to contribute to the field of study, an algorithm based on decomposition that progressively adapts its weight vectors during the evolution process using a archive of nondominated solutions is proposed. The proposed algorithm is called Multi-objective Evolutionary Algorithm based on Decomposition with Local-Neighborhood Adaptation (MOEA/D-LNA). Subsequently, the proposed algorithm is compared to other algorithms from the literature in three sets of test functions, DTLZ, WFG, MaF and the one resulting from this research Generalized Position-Distance (GPD), with different weight vector initialization procedures in 3,5,8 and 10 objectives. The results have shown interesting characteristics and promising results on irregular Pareto fronts.For example on the problems DTLZ5. IDTLZ1, MaF1, GPD1 e GPD2.