スケジュール化された局所重み付き回帰によるフィードフォワード学習制御

Abstract

Feedback Error Learning (FEL) has been actively studied in an adaptive control framework. When the plant is a time-variant or nonlinear system, however, conventional FEL schemes do not work any more because a single inverse model is not enough. As a remedy to this problem, Just-In-Time (JIT) modeling has been proposed, where we memorize regression data for all input/output behavior of the plant and construct local models as required in real time. However, a scheme for implementing JIT modeling, Locally Weighted Regression (LWR), has problems that it is computationally heavy to search for query point among huge amount of the past input/output data, and that regression calculation is difficult when the number of information vectors is insufficient. In view of these points, this paper proposes a new scheme. In the proposed scheme, we seek local models by a schedule parameter, instead of searching time series by input/output data. We call this scheme Scheduled LWR. By applying it to the FEL, we can reduce the search time without losing practicality and also solve the lack of sample number. We verify the effectiveness of the proposed scheme by simulation using a 2-link robot arm model.