A learning algorithm to select consistent reactions to human movements

Abstract

A balance between adaptiveness and consistency is desired for a robot to select control laws to generate reactions to human movements. Two existing algorithms, the weighted majority algorithm and the online Winnow algorithm, are biased for either strong adaptiveness or strong consistency. The dual expert algorithm (DEA), proposed in this paper, is able to achieve a tradeoff between consistency and adaptiveness. We give theoretical analysis to rigorously characterize the performance of DEA. Both simulation results and experimental data are demonstrated to confirm that DEA enables a robot to learn the preferred control law to pass a human subject in a hallway setting.