Abstract:
Noise rejection, handling the plant-model mismatches and alleviation of structured or unstructured uncertainties constitute prime challenges that are frequently encountered in the practice of systems and control engineering. One way of reducing the adverse effects of the stated difficulties and obtaining a good tracking precision is to utilize the techniques of variable structure systems theory, which offers well formulated solutions particularly to problems containing uncertainty and imprecision.In this thesis, variable structure systems theory based training strategies of computationally intelligent systems are discussed. Two approaches are developed for alleviating the above mentioned difficulties. Additionally, the learning rate selection problem is treated from the point of variable structure control.In the first approach described, a dynamic parameter adaptation law is derived and the applicability of the algorithm is discussed. The analysis presented aims to extract the conditions for establishing equivalence between sliding mode control of the plant and sliding mode learning in the controller. The second method is based on the selection of an extended Lyapunov function, by the use of which the sensitivity of the cost measure to the adjustable parameters are minimized together with the half squared error measure. Lastly the selection of the learning rate for three different gradient based parameter tuning strategies are discussed. The objective of the learning rate selection is to drive the plant to a sliding mode while the output of the controller is driven to a similar regime.The performances of the methods developed are assessed on the dynamic model of a two degrees of freedom direct drive SCARA robotic manipulator, whose dynamic equations are assumed to be unknown throughout the results presented. In the simulations, the alleviation of the adverse effects of observation noise and varying payload conditions are studied.
