Two suggestions for efficient implementation of CMAC's

Abstract

The CMAC (Cerebellar Model Articulation Controller), first introduced by Albus [2] is a widely known and used neural network model. with applications in robotics. control. digital communications. and many others. It is normally employed as a method for function approximation. whose convergence using a well defined simple learning algorithm has been proved [4]. Also. its modeling capabilities have been characterized [3]. Nevertheless. it suffers from two important problems: the huge amount of memory needed for its implementation in many common situations. and the lack of a systematic way for selecting appropriate values for its parameters.particularly number of quantization intervals. This communication presents two proposals for addressing these difficulties: the first one is a dynamic implementation that books memory only for those weights needed to represent the training data set. and that performs linear interpolation when a query using other weights is requested. The second one consists on the definition of an index of correlation from which the optimal number of quantization intervals that should be assigned to each dimension of the input space can be found. Experiments are performed for two synthetic cases and for one set of real data. These are used to model the dynamic behaviour of a real sensorized car. Figures are given to show the memory savings and mean squared error obtained.

Hierarchically Clustered Adaptive Quantization CMAC and Its Learning Convergence

“…The difference of the HCAQ-CMAC memory contents between two successive iterations for the th input training sample (denoted as ) is, therefore, defined as (23) Note that the activation mask is a constant for an arbitrary input training sample across different training iterations. This is because the HCAQ-CMAC network structure is static after the structural learning phase.…”