Multiple radial basis function networks in modeling and control

A blending of radial basis function networks and general regression neural networks is proposed and shown to be suitable for the modeling and control of dynamical systems. In this blended scheme, the radial basis function networks account for long term and global dynamics, while the general regression neural networks govern short term and local dynamics. Due to the similarity between radial basis function networks and general regression neural networks, which are both synthesized using Gaussian nodes, the blended scheme can be congruently implemented as a multiple radial basis function network. Training techniques of the two constituting networks are explored in the multiple radial basis function networks so as to achieve a balance in temporal as well as spatial characterization. This leads to a high precision, highly dynamical neural network system. Applications of this multiple radial basis function network in aircraft modeling and control are conducted. Real-time training and precision enhancement are demonstrated.