Insertion of Artificial Individuals to Increase the Diversity of Multiobjective Evolutionary Algorithms
Jamisson Santana Freitas, Cícero Garrozi, Mêuser Valença · 2015
In recent years, many evolutionary algorithms approaches were introduced to improve existing algorithms as well as to solve optimization and search for problems. Problems involving the optimization of many objectives require a set of optimal solutions known as Pareto frontier. Unfortunately, similarly to single objective Evolutionary Algorithms, the Multiobjective Evolutionary Algorithms (MOEAs) also suffer from loss of genetic diversity. When solutions converge to a few Pareto points, a mechanism to maintain the diversity of the population throughout generations is needed. It is expected that, if diversity is controlled effectively, at the end of the evolutionary process, the Pareto Front optimum will be distributed as uniformly as possible. This paper proposes a new diversity operator which generates artificial solutions to fill sparse regions of the non-dominated set of solutions found by the MOEA. It uses Artificial Neural Networks (ANN) to perform a reverse mapping from the phenotype to the corresponding genotype of an inserted artificial solution; this mechanism was tested with NSGA-II and SPEA2 algorithms. With the diversity operator is possible to obtain significant improvements in the hypervolume metric and, mainly, a reduce in the spread metric in the solutions of the Pareto frontier obtained.