Support vector regression based controller design methods for nonlinear systems

Conventional nonadaptive controllers may be convenient enough for most industrial processes in the case that the system does not include highly nonlinear and time varying dynamics. Depending on the developments in technology and increasing requirements of the people, systems are transforming into more and more complex structures. Nonadaptive controllers cannot counteract the tracking error in case the behaviour of the system changes perceptibly owing to internal or external factors during the course of operation and the control performance degrades. For this purpose, different controllers for different scenario conditions and operating points may be designed to resume successful control in industry. Controller design depending on different performance criteria is a momentous task since effective design of controllers can significantly reduce the costs in industry. However, that chore can also be automatically executed by an adaptive controller. Therefore, controllers that are able to adapt themselves according to varying system dynamics must be developed using optimization theory based solutions to enhance the success of the controllers in industry. Adaptive controllers can generally be classified as model based adaptive controllers and model free adaptive controllers. Model based adaptive controllers utilize the system model to approximate the behaviour of the system dynamics whereas in model free adaptive controllers the adjustment mechanism does not need the model of the system. The performances of model based adaptive controllers are directly influenced by the accuracy of the system model. Machine learning algorithms have frequently been utilized to identify the dynamical behaviour of the system precisely in order to derive effective adjustment rules for the parameters of adaptive controllers. Support Vector Regression (SVR), proposed by Vladimir Vapnik et al., is one of the most favorable nonlinear system identification methods since it guarantees the global extremum of the optimization problem. Various adaptive controllers based on SVR have been proposed for the control of nonlinear systems in technical literature. The common property of these studies is that SVR is typically deployed to approximate the system Jacobian which is required to tune parameters of adaptive controllers through gradient based optimization algorithms. In this thesis, three novel adaptive controller architectures based on SVR have been proposed.The basic novelty in these architectures is that SVR has been implemented directly as a controller and a parameter estimator for a generalized controller structure. The performance evaluations of the proposed controllers have been examined on various nonlinear benchmark systems by simulations. In addition, stability analysis of the systems have been performed. The obtained results prove that the proposed adjustment mechanisms are effective in controlling processes with nonlinear dynamics, noise and uncertainties.