Training of fuzzy neural networks via quantum-behaved particle swarm optimization and rival penalized competitive learning

There are some difficulties encountered in the application of fuzzy Radial Basis Function (RBF) neural network. One of them is how to determine the number of hidden rule neurons and another difficulty is about interpretability. In order to overcome these difficulties, we have proposed a fuzzy neural network based on RBF network and takagi-sugeno fuzzy system. We have used a new structure of fuzzy RBF neural network, which has been proved that it is better than other structures in term of interpretability. Our model also use a Rival Penalized Competitive Learning (RPCL) and a swarm based algorithm called Quantum-behaved Particle Swarm Optimization (QPSO) to determine design parameters of hidden layer and design parameters of output layer, respectively. RPCL is the best clustering algorithm that is introduced so far. The Particle Swarm Optimization (PSO) is a well-known population-based swarm intelligence algorithm. The QPSO is also proposed by combining the classical CPSO philosophy and quantum mechanics to improve performance of PSO. We have compared the performance of the proposed method with gradient based method. Simulation results of nonlinear function approximation demonstrate the superiority of the proposed method over gradient based method.