On the convergence of the Albus Perceptron

Abstract

Albus (1975a,b, 1981) developed CMAC, an adaptive system for robotic control, based on the cerebellum and the classical perceptron. He applied it to controlling a physical model of the human arm, with seven degrees of freedom. The system exhibited the classical learning curve, generalization, and learning interference. This paper proves the convergence of this learning scheme. Computer experiments are presented investigating the effects of training strategy and generalization width on the rate of convergence and on learning interference. 1.