A fuzzy control based algorithm to train perceptrons

In this paper a method to train perceptrons using a fuzzy controller is presented. When the first layer of a perceptron is trained, the fuzzy rules try for each connection of a neuron that the weight is similar to the input of the connection if the desired output of the neuron is high, otherwise the fuzzy rules try the one that the weight is different to the input of the connection. When the rest of the connections of a perceptron are trained, the fuzzy rules try, besides modifying the weights, to return the desired outputs for the neurons of the previous layer in the perceptron. The training of multilayer perceptrons with neurons whose activation function is not differentiable has been attained with this method