A Multi-Objective Evolutionary Algorithm using Neural Networks to Approximate Fitness Evaluations

Abstract

Two different methods to accelerate the search of a Multi-Objective Evolutionary Algorithm (MOEA) using Artificial Neural Networks are presented. Two different methods are proposed. One using ANN to approximate the fitness of the solutions alternated with the real fitness evaluation, being the ANN approximation used only when the estimated error of the neural network was lower than a pre-defined value. In the second method, the ANN is used as a local search strategy by defining new better solutions from the precedent generation. These methods can substantially reduce the number of fitness evaluations on computational expensive problems while not compromise the good search capabilities of MOEA. The efficiency of the methods proposed is tested on several benchmark functions as well on a real multi-optimization problem of polymer extrusion.