A direct model reference approach for implementation of neurocontrollers

For a special class of single-input single-output (SISO) nonlinear systems, a direct method is presented to learn the control law-by a multilayer network-such that the overall behavior of the closed-loop system will be the same as that of a reference model. The method is inspired by a direct adaptive control scheme for direct adjustment of the weights of a multilayer network implementing the controller using backpropagation. To guarantee correct learning of the control law, the states of the model must be kept close to those of the nonlinear plant. This requires the addition of the resetting scheme to the overall architecture. The linear model states are reset to the states of the nonlinear plant at the end of a resetting time-window of length m and the error measured at the end of the resetting window is used for weight adjustment. The length of the time-window is kept small relative to the dominant dynamics of the plant