Abordagens multi-objetivo para o treinamento de redes neurais e seleção de características

Abstract

Artificial neural networks have been successfully applied in solvingproblems such as functions approximation and patterns classification,where the extraction of a model can be difficult to see. The search for a model that best represents the problem makes the generalizationability the main concern in the training of artificial neural networks, atask that becomes even more difficult in environments with large dimensionality. In this context, this paper proposes new techniques for training multiobjective neural network, where the minimization of the risk and the control of complexity are objectives to be achieved through of the training in order to obtain a model more compatible to the problem. It also proposed an approach to dimensionality reduction through the task of feature selection, in which the objectives are to reduce the number of attributes of the problem and maximize the correct classification rate, making it less arduous task of classifiers in environments with large numbers of dimensions.