An introspective algorithm for achieving low-gain high-performance robust neural-adaptive control

Abstract

A method proposed for halting weight drift in neural-adaptive control schemes is analyzed using the method of describing functions. The method utilizes a self-evaluating, introspective method with a Cerebellar Model Arithmetic Computer. The average error within the domain of local basis functions is measured, and then used to estimate the effect of weight updates on reducing the error i.e. estimating a partial derivative. The adaptation algorithm halts the weight updates when it is determined that weight updates are no longer beneficial in reducing the average error. In this paper, a describing function analysis establishes stability assuming an accurate measure of the partial derivative.