Fuzzy logic controlled multilayer neural networks: theory and case studies

A fuzzy logic controlled learning algorithm for training multilayer feedforward networks (FCNN) is presented. Two fuzzy controllers, the /spl eta/(t)-fuzzy controller and the /spl alpha/(t)-fuzzy controller, are proposed to adaptively adjust the learning rate /spl eta/(t) and the momentum coefficient /spl alpha/(t) during the training process. An electric power transformer initial fault diagnosis and the XOR problem are taken as examples to test the performance of the proposed FCNN. It is demonstrated that the proposed FCNN has much better performance than Back-propagation (BP) algorithms and simulated annealing algorithms. The integration of various intelligence techniques to form a hybrid system such as the FCNN is a very important way forward in the next generation of intelligent system design.