Rede Neural Difusa com T-normas Diferenciáveis e Interativas

AUTOR(ES)
DATA DE PUBLICAÇÃO

2007

RESUMO

Fuzzy sets are used in the representation of vague and imprecise knowledge. Neural networks, besides their computational parallelism, also have learning capabilities. The combination of such both paradigms is an attempt to congregate their benefits in an integrated system, such a fuzzy neural network. T-norms are functions that actuate like intersection and union operators. Most of times, fuzzy neural networks use the minimum and the maximum t-norms to perform the conjunction and disjunction, respectively. The use of such t-norms makes impossible the direct application of the backpropagation algorithm, based on the descendent gradient method, given the absence of the mean square error derivative. A neural network, most of times, does not allow the extraction of the knowledge encoded into its synaptical weights. In other words, it is not capable to make more explicit its decision reasoning. Despite this, there are several rule generation algorithms from trained neural networks. The number of rules generated by these algorithms might be very large, making hard to analise them. It has presented various evaluation measures and post-processing algorithms that allows the user to focalize his attention to the better ones in the considered rule set. The main goal of this project is to apply differentiable and interactive t-norms to perform the conjunction and disjunction operations in a fuzzy neural network, to pattern classification and rule generation. The differentiable fuzzy neural network, denominated RNDD, object of this dissertation, is a feedforward fully-connected network with three layers: the input layer, the hidden layer of and neurons and the output layer of or neurons. It was used the backpropagation algorithm for training. Rules will be generated by the backtracking algorithm because of its intuitive nature an to be applicable to RNDD. The proposed methodology was evaluated over an application in the Iris and Vowel datasets and other ones randomly generated with promising training, test and rule generation results. For the classification tests over Vowel dataset, the system reached a hit rate of 80.3%. The Kappa coefficient was calculated and the value reached was 0.771. From the trained network with the Iris dataset, it was generated a set of fuzzy rules justifying the network decision. After the development of the proposed system, it be expected to obtain an automatic method of knowledge acquisition from data examples.

ASSUNTO(S)

geração de regras sistema neuro-fuzzy ciencia da computacao

ACESSO AO ARTIGO

Documentos Relacionados