Comparison of convergence properties of CMAC neural networks and traditional adaptive controllers

The self-tuning regulator, the model-reference adaptive method, and the cerebellar model articulation controller (CMAC) neural network method are compared for speed of adaptation, robustness to model mismatch, tracking error, control effort and global stability. The three control methods are analyzed for the same problem. Results indicate that the CMAC approach exhibits desirable properties not found in the other two methods. The CMAC method works well for nonlinear systems, performs well in noise, and can be implemented efficiently for large-scale systems. However, the CMAC method may be stable only for some memory update rates and generalization factors