Neural networks composed of single-variable CMACs

This paper presents a CMAC-based neural network that needs much smaller memory space compared to the conventional CMAC. The used neural network has a modulated structure composed of single-variable CMAC. CMAC is a table look-up neurocomputing technique capable of learning static mapping. However, it suffers from the "curse of dimensionality". Using only single-variable CMAC in neural network significantly reduces the needed memory space and overcomes the enormous memory size problem in the conventional CMAC in high-dimensional modeling. With the same size of memory, the new structure is able to achieve much smaller learning error compared to the conventional CMAC. With the modularity of the neural network structure, the learning can be decomposed into several stages. A neural network with an initial number of modules is used to learn primary skills. To develop more advanced techniques, one or more modules is added to the network. Attractive features of the new learning scheme include modular structure, system expansibility, and potential faster learning.