An evolutionary RBF networks based on RPCL and its application in fault diagnosis

The performance of a RBF neural network strongly depends on the network structure and parameters. Therefore, an algorithm that can automatically select the network configuration will be very beneficial. This paper presents a novel method where configuration of an RBF network can be learned by a hybrid training schema. Appropriate number of neural nets in the hidden layer and their coarse centers is obtained by a modified version of Rival Penalized Competitive Learning (RPCL). Then, Genetic Algorithm is applied to determine the optimal parameters including refined cluster centers, variance of Radial-Basis Function and weights. The application of proposed evolutionary RPCL-RBF networks is discussed and the result of simulation illustrate that proposed method is effective.