Polibits, Vol. 46, pp. 13-27, 2012.
Abstract: Artificial Neural Networks (ANN) have been applied in several tasks in the field of Artificial Intelligence. Despite their decline and its resurgence, the ANN design is a testing-error process which can stay trapped in bad solutions. In addition, the learning algorithms used like back-propagation and others based in the descendent gradient present a disadvantage: they cannot be used to solve non continuous and multimodal problems. For this reason, the application of evolutionary algorithms to automatically design ANN is proposed. In this research, the Differential Evolution (DE) algorithm finds the best ANN principal elements: the architecture, the set of synaptic weights and the set of transfer functions. Also two fitness functions are used (the mean square error - MSE and the classification error - CER) which involve the validation stage to guarantee a good ANN performance. First, a study of the best parameter configuration for DE algorithm is done. The experimental results show the performance of the proposed methodology to solve pattern classification problems. Also, a comparison with two classic learning algorithms: descendient gradiant and Levenberg-Marquardt is presented.
Keywords: Evolución diferencial, evolución de redes neuronales artificiales, clasificación de patrones
PDF: Diseño automático de redes neuronales artificiales mediante el uso del algoritmo de evolución diferencial (ED)
PDF: Diseño automático de redes neuronales artificiales mediante el uso del algoritmo de evolución diferencial (ED)