Generalization Ability and Learning Accuracy of Hyperball CMAC and Parallel CMAC Structure
Pei-Yong Duan,Hui-He Shao
DOI: https://doi.org/10.1016/s1474-6670(17)56890-1
1999-01-01
IFAC Proceedings Volumes
Abstract:A cerebellar model articulation controller (CMAC) is often used as nonlinear function approximator in pattern recognition, system identification, and neurocontrol systems. Generalization ability and learning accuracy are among important performance of a CMAC besides learning speed, and they are affected significantly by quantization level and the ratio of adjusted weights at each learning with a sample to all weights, called weigh adjusting ratio (WAR) in this paper. There is always a trade-off between generalizatinn ability and learning accuracy in conventional neural networks. In many cases, quantization level is increased to enhance learning accuracy without consideration of changing the WAR, this may result in decreasing generalization ability. In this paper, we studied Hyperball CMAC (HCMAC), and found the relation of generalization ability or learning accuracy with quantization level and WAR, a large number simulations illustrate that the CMAC is very effective in approximating nonlinear functions. On one hand, for fixed quantization level, the generalization error and learning error will conformably decrease with increasing WAR until WAR reaches a certain value, and they remain unchanged when the WAR is greater than the value. On the other hand, when the WAR fixed. learning error is decreasing with large quantization level, whereas generalization error has no clear relation with quantization level. These unique properties can be applied to improve the performance of designed CMAC. For further enhancing generalization ability and learning accuracy, a multiple CMAC structure (called parallel CMAC structure in this paper) is proposed based on the CMAC to approximate a highly nonlinear function very accurately. A simulation example demonstrates that the structure has powerful generalization and learning ability