Comparing Evolutionary Artificial Neural Networks from Second and Third Generations for Solving Supervised Classification Problems

Abstract

Constituting nature-inspired computational systems, Artificial Neural Networks (ANNs) are generally classified into several generations depending on the features and capabilities of their neuron models. As generations develop, newer models of ANNs portrait more plausible properties than their predecessors, accounting for closer resemblance to biological neurons or for augmentations in their problem-solving abilities. Evolutionary Artificial Neural Networks (EANNs) is a paradigm to design ANNs involving Evolutionary Algorithms (EAs) to determine inherent aspects of the networks such as topology or parameterization, while prescinding—totally or partially—from expert proficiency. In this paper a comparison of the performance of evolutionary-designed ANNs from the second and third generations is made. An EA-based technique known as Grammatical Evolution (GE) is used to automatically design ANNs for solving supervised classification problems. Partially-connected three-layered feedforward topologies and synaptic connections for both types of considered ANNs are determined by the evolutionary process of GE; an explicit training task is not necessary. The proposed framework was tested on several well-known benchmark datasets, providing relevant and consistent results; accuracies exhibited by third-generation ANNs matched or bested those from second-generation ANNs. Furthermore, produced networks achieved a considerable reduction in the amount of existing synapses, as in comparison with equivalent fully-connected topologies, and a lower usage of traits from the input vector.