Constructive and Pruning Methods for Neural Network Design

Marcelo Azevedo CostaAntônio de Pádua BragaBenjamin Rodrigues de Menezes

This paper presents methods to improve generalization of Multi-Layer Perceptron (MLP) by pruning the original topology without loss in performance. Topology information and validation sets are used. The results show that these techniques are able to choose a minimum network topology and to simplify trained networks.

Caso o link acima esteja inválido, faça uma busca pelo texto completo na Web: Buscar na Web

Biblioteca Digital Brasileira de Computação - Contato:
     Mantida por: