Neural network-based identification and adaptive control of nonlinear systems: a novel dynamical network architecture and training policy

The authors present a novel dynamical neural network structure and a training algorithm for the identification and adaptive control of nonlinear systems. The multilayer network, which consists of a hidden layer of dynamical nodes with recurrent connections and the supervised training scheme that employs an LMS updating rule for the interconnection weights, provides significantly improved performance compared to the standard propagation networks used in these applications. A specific application of the controller design approach is outlined for the online identification and decentralized adaptive control of multijointed robotic manipulators.