Adaptive Control in the Presence of Outliers

Abstract

This paper presents a modification of a predictive adaptive controller that renders it robust with respect to the occurrence of outliers in the plant measured output. Outliers are large deviations of the signal being measured that are not explained by a Gaussian distribution. Making a Gaussian assumption on the statistics of the observation noise results in the use of a quadratic loss for designing either estimators or controllers. In turn of a quadratic loss yields major amplification of large signal deviations causing poor parameter estimates and, consequently, controller gain detuning and loss of performance. The algorithm presented to solve this problem relies on a modification of the cost being minimized, such as to render it more insensitive to large prediction error values. Simulations on the position control in a ball and beam plant are presented to illustrate the advantages of the controller proposed