Direct neural model reference adaptive control

Abstract

The paper investigates in detail the possible application of neural networks to direct model reference adaptive control. The difficulties involved in training the neural controller embedded within the closed loop are discussed in detail. A training structure is suggested that removes the need for a generalised learning phase. Techniques are discussed for the backpropagation of errors through the plant to the controller. In particular, dynamic plant Jacobian modelling is proposed that uses a parallel neural forward model of the plant. The benefits of neural control are then demonstrated by comparison with Lyapunov adaptive control for a number of example plants. A continuously stirred tank reactor and a nonlinear guidance system are chosen as two realistic nonlinear case studies for the demonstration of the techniques discussed. In both cases nonlinear neural control was found to provide greatly improved performance over conventional approaches.