A comparison of a neural network and a model reference adaptive controller

A two-mode coupled compound pendulum is used to compare a computed-torque-type model reference adaptive controller and a feedforward neural network controller. A derived globally asymptotically stable adaptation law for the neural net controller shows that the back error propagation scheme used is, in some cases, also asymptotically stable. Computer simulations of the two controllers demonstrate their relative performance. This comparison shows that the derived adaptation law compares favorably with the performance of the model reference adaptive controller. It also lends insight into the required input signal frequency content in order to guarantee proper convergence of the neural network. The convergence and stability properties of the neural network when it is used as a feedforward computed-torque controller are analyzed.<>