Fuzzy logic adapted nodal training parameter
M.S. Gelder · 2002
A technique is outlined for improving the learning rate of a multilayer perceptron (MLP) network. Each network node is assigned its own training rate parameter which is adapted using fuzzy logic as part of the error backpropagation process. This involves the development of target values for hidden layer node output. These values are based on the current network weight state and are therefore different for each epoch. Using two test vector distributions it is demonstrated that this approach can reduce MLP convergence time and is compared to three other training methods: standard backpropagation, fuzzy adapted global training rate parameter, and the delta-bar-delta learning rule.