A stochastic method for neural-adaptive control of multi-modal nonlinear systems

The multiple model adaptive control approach is extended to a class of nonlinear stochastic systems whose underlying functions are unknown and which can change arbitrarily in time. Gaussian radial basis function neural networks are used to learn the nonlinear functions characterising the different plant modes online, without resorting to a separate learning phase. Function estimation, mode change detection and control signal generation are based on probabilistic techniques utilising concepts of Kalman filtering, the multiple model algorithm and dual control.