Least mean square learning in associative memory networks

The authors investigate theoretically the use of a class of neural networks called associative memory networks for online adaptive nonlinear modeling and control. This class of networks is defined to include such algorithms as the cerebellar model articulation controller (CMAC), B-splines, and fuzzy logic. The algorithms are defined within a unifying framework that provides a natural decomposition for a parallel implementation. The modeling capabilities of the networks are investigated, and some new results on the CMAC are presented. The instantaneous learning rules are derived and investigated from a geometrical perspective, which allows the rate of convergence to be analyzed. A measure of the learning interference for different set shapes is obtained. An example of an associative memory network guiding an autonomous vehicle into a slot is given