Local Online Support Vector Regression for Learning Control

Abstract

Support vector regression (SVR) is a class of machine learning technique that has been successfully applied to low-level learning control in robotics. Because of the large amount of computation required by SVR, however, most studies have used a batch mode. Although a recently developed online form of SVR shows faster learning performance than batch SVR, the amount of computation required by online SVR prevent its use in real-time robot learning control, which requires short sampling time. Here, we present a novel method, Local online SVR for Learning control, or LoSVR, that extends online SVR with a windowing method. We demonstrate the performance of LoSVR in learning the inverse dynamics of both a simulated two-joint robot and a real one-link robot arm. Our results show that, in both cases, LoSVR can learn the inverse dynamics on-line faster and with a better accuracy than batch SVR.