Stable neural control of a flexible-joint manipulator subjected to sinusoidal disturbance

Abstract

The proposed method aims at halting weight drift when using multilayer perception backpropagation networks in direct adaptive control schemes, without sacrificing performance or requiring unrealistic large control gains. Unchecked weight drift can lead to a chattering control signal and cause bursting. Previously proposed robust weight update methods, including e-modification and dead-zone, will sacrifice significant performance if large control gains cannot be applied. In this work, a set of alternate weights guides the training in order to prevent drift. Experiments with a two-link flexible-joint robot demonstrate the improvement in performance compared to e-modification and dead-zone.