Realisation of neural network controllers in integrated process supervision

Recent developments in neural network controllers have focused mainly on either primary or adaptive control techniques. To date, there has been little attempt to integrate and schedule them within a common control framework on the basis of the system behaviour. An architecture for integrated process supervision, IPS, has been proposed by Leitch and Quek (IEE Proc.-D, Control Theory and Application, 39(3) (1992) 317-27) as a general meta-level supervisory system which automatically schedules between generic control tasks according to the system performance. The IPS scheme was successfully validated using various classical and adaptive controllers (Leitch and Quek, IEE Proc. 3rd Int. Conf. Control, Vol. 1, March, 1991, pp. 127-33; Ho and Goh, Final Year Dissertation, Nanyang Technological University, 1993). This paper demonstrates how the IPS scheme can be used to integrate and schedule between the neural network primary and adaptive control regimes. The cerebellar model articulation controller (Conforth & Elliman, in Techniques and Applications of Neural Networks, ed. M. Taylor & P. Lisboa. Prentice Hall, UK, 1993, pp. 35–46), CMAC, is chosen for this purpose. Its structure is modified and integrated within the IPS scheme. The modification results in better system performances than the standard PI controllers. Moreover the realisation of the modified CMAC and its associated learning algorithm within the IPS illustrates the generality of the IPS as a generic meta-level supervisory control architecture.