Model free adaptive support vector regressor controller for nonlinear systems

• Published in: Engineering applications of artificial intelligence 2019-05, Vol.81, p.47-67
• Main Authors: Uçak, Kemal, Günel, Gülay Öke
• Format: Article
• Language: English
• Subjects: Adaptive control; Direct adaptive control; Model free adaptive control(MFAC); Model free SVR controller; Online support vector regression; SVR controller
• ISSN: 0952-1976; 1873-6769
• DOI: 10.1016/j.engappai.2019.02.001

In this study, a novel model free support vector regressor controller (MF-SVRcontroller) is introduced for nonlinear dynamical systems. For the adaptation mechanism, a model free closed-loop margin which is a function of tracking error is derived and it is used to optimize the parameters of MF-SVRcontroller. The effectiveness of the adjustment mechanism and closed-loop performance of the MF-SVRcontroller have been examined by simulations performed on continuously stirred tank reactor (CSTR) and bioreactor benchmark systems. In order to observe the impacts of the removal of the model estimation block in control architecture, the performance of the MF-SVRcontroller is compared with a model based support vector regressor controller (MB-SVRcontroller) and SVM-based PID controller. The results indicate that MF-SVRcontroller diminishes the computational load of MB-SVRcontroller at the cost of a small amount of decrease in tracking performance. •The gradient effects which occur in NN and ANFIS are vanished in SVR and global extremum is ensured.•In SVR, a non-convex optimization problem in primal form is tranformed to a new form called as dual form which is a convex objective function with linear constraints.•A novel model free adaptive SVR controller (MF-SVRcontroller) is proposed to directly control nonlinear dynamical systems.•The main contribution of MF-SVRcontroller is adjusting the parameters of SVR controller via tracking error without using the system model.•The most significant strength of MF-SVRcontroller is that the structure does not require any system identification phase.