Neural network synthesis

Abstract

This report describes a feed forward Artificial Neural Network (ANN) synthesis via an Analytic Programming (AP) by means of the ANN creation, learning and optimization. This process encompasses four different fields: Evolutionary Algorithms, Symbolic Regression, ANN and parallel computing to successfully synthesize a suitable ANN within a reasonable time. ANN synthesis proved to be a useful and efficient tool for nonlinear modeling and its results were applied to intelligent system controlling an energetic framework of an urban agglomeration. Furthermore, the ANN synthesis proved to have the ability to synthesize smaller ANN than the Genetic Programming (GP) while simultaneously almost infinitely complex ANN can be synthesized by the application of multiple evolution loops. This process can also produce ANN with feed forward branching, which is an unavailable quality for the GP.