Authors: Felipe Trujillo-Romero
Addresses: Division de Estudios de Posgrado, Universidad Tecnologica de la Mixteca, Carretera a Acatlima Km. 2.5 Huajuapan de Leon, Oaxaca, Mexico
Abstract: This paper discusses the generation of neural networks that are obtained from the evolution of individual's population in a genetic algorithm. For achieving this, the population of individuals for the genetic algorithm is formed of structural elements which constitute the neural networks. These elements include the number of layers, neurons per layer, transfer functions and the connections between neurons in the network, among others. These individuals as can be seen a structure which has the ability to evolve rather than a standard genotype. Furthermore, the size of the individuals is not defined and depends mainly on the neural network which in turn depends on the problem to be solved. This structure considered as an evolutionary entity, is able to evolve until convergence towards a suitable structure is achieved. The fitness function is specified with the features of the problem to be solved by the neural network. This algorithm has been tested successfully in solving classification problems, as in the case of alpha-numerical character recognition, and has been compared against a neural network obtained by conventional means. Better results were obtained with the neural network generated by using genetic programming of this type of evolutionary entities.
Keywords: neural networks; evolutionary algorithms; evolutionary entity; genetic algorithms; alpha-numeric character recognition; classification.
International Journal of Bio-Inspired Computation, 2013 Vol.5 No.5, pp.289 - 302
Available online: 16 Oct 2013 *Full-text access for editors Access for subscribers Purchase this article Comment on this article