A robust growing-pruning algorithm using fuzzy logic

We describe a fuzzy-logic controlled growing and pruning algorithm for improved fault-tolerance in a multilayer perceptron. This algorithm is intended for classification tasks. The importance of each weight connection, and hence of each hidden neuron, of the network is determined as a fuzzy-logic function of the increase in the mean-squared error and the decrease in the number of correct classifications upon the injection of faults. The hidden neurons may be either duplicated or pruned depending upon their relative importance in the classification task. We compare our network to a 4-augmented multilayer perceptron with 12 hidden neurons and find that they both have the same level of fault tolerance, however the former is smaller. We then use the immunization technique with our algorithm and find that the number of weight connections can be further significantly reduced