|
The Albus's Cerebellar Model Articulation Controller (CMAC) has
been used in many practical areas with considerable success and
capable of learning nonlinear functions extremely quickly due
to the local nature in weight updating. Besides, the higher
-order CMAC model proposed by Stephen and David adopts B-Spline
receptive field functions and a more general addressing scheme
for weight retrieving, which can learn both functions and func-
tion derivatives. In this thesis, we present a three-layered
fuzzy CMAC network, which takes the bell-shape membership func-
tions as the receptive field functions and use the centroid of
area(COA) approach as the defuzzification interface. The learn-
ing algorithm is based on the maximum gradient method. For the
situation of insufficient and irregularly distributed training
patterns, we propose a sampling method based on interpolation
scheme to generate the proper training patterns. The proposed
fuzzy CMAC model is basically a table look-up model in whih
fuzzy weights are stored and manipulated by using fuzzy set
theory. This model adaptively adjusts the weights according to
sample data to approximate the nonlinear continuous functions.
Finally, we take some experiments including gerneral function
approximation and color correction to verify the proposed model.
|