A general learning scheme for CMAC-based controller

Cerebellar model articulation controller (CMAC) is a powerful tool for nonlinear control applications. However, it yet lacks an adequate learning scheme. It is found that, with the existing learning scheme, if a complicated learning algorithm is not used, CMAC can destabilize a system that is otherwise stable. Oscillations resulting from the interaction between CMAC and the classical controller were found to contribute to the instability. This paper presents a new CMAC learning scheme that models plant's characteristics based on closed loop errors instead of the original input-output pairs. In this scheme, memory space of the CMAC is partitioned into two parts. One is for dynamic control, in which dynamic information is stored. Another is for steady state control, in which steady state information is adaptively updated for smooth control. Relationship between the two parts of the space is discussed and specified for a stable control. Simulation results on a typical nonlinear plant model and a real electrohydraulic servo system using the proposed scheme demonstrate that the oscillations are eliminated and stable control is obtained. The new scheme demonstrates superior tracking performance, noise rejection property and good robustness.