Feedforward Learning Control by Scheduled Locally Weighted Regression

It is widely recognized that feedforward (FF) control is effective for response shaping but requires an accurate model of the plant. In view of this, on-line FF learning control has been studied extensively in literature. This paper proposes a learning scheme for FF control based on scheduled locally weighted regression (LWR). When the plant dynamics change with a schedule parameter available in real time, the input/output behavior is labelled by the parameter and accumulated as regression data. FF control signal is then generated by means of regression weighted locally with respect to the current schedule parameter (query). The proposed scheme is illustrated by numerical simulation of i) a linear time-varying system and then ii) a two-link robot arm which is highly nonlinear and challenging.