Error prediction and structure determination for CMAC neural network based on the uniform design method

Insufficient study on error bound of cerebellar model articulation controller (CMAC) severely limits its application. To investigate the error prediction and structure determination of CMAC for multi-dimensional and data-generation objects, this paper builds a 10-input 2-output model for a desulfurization system to test 44,640 sets of operation data. Four test groups and one prediction group are designed and performed based on uniform design method. Regression analysis and curve fitting methods are applied for error analyses. The focus of regression analysis method is the influence of uniform table's level on its prediction formulas' accuracies, whereas curve fitting's is the impact of theoretical memory space (location number) and actual storage space (address number) of CMAC on output error. Based on the results, the following conclusions are obtained. (a) The prediction accuracy of the linear regression equation is not monotonous with the level of the uniform design table, but there is a distinct region with local high precision. (b) Compared with regression analysis and address number fitting methods, location number analysis method has distinct advantages of prediction range, accuracy and flexibility. (c) In terms of location number analysis, different intervals may correspond to different optimal fitting functions, but only power function maintains high prediction accuracy in the whole range where the method works. Besides, the scope of location number analysis is also studied, of which the lower borders are 10⁹ and 10¹⁰ approximately for model's two outputs, respectively.