Hyperball Cerebellar Model Articulation Controller (CMAC) Based System Modeling and Its Modifying Strategies

Hyperball cerebellar model articulation controller (HCMAC) has advantages such as simple structure, fast learning convergence and powerful generalization capability. In order to overcome the shortcomings of the conventional neural networks based modeling methods, the HCMAC based modeling methods of nonlinear dynamic systems were presented. The modifying error ratio based multiple-stepped on-line modifying strategy was given according to the qualitatively analyzed causes leading to modeling errors. In order to exactly gain the dynamic performance changed with a variant operation point of a certain industry process, the model expands its learning sample data space pertinent to the variation to gain more information near the operation point. In addition, part of each error between real output of the plant and the estimated by network model is applied to iteratively train the model on-line. The simulation results demonstrate that the methods are capable of decreasing the initial model errors resulting from the insufficient or/and inaccurate learning sample data, and the model can be easily and on-line adjusted to follow a practical varied operation point of the nonlinear dynamic process.