強化学習を用いた大車輪ロボットの学習知識の選択可能性について

概要

論文の詳細を見る
We have studied a giant swing robot whose giant swing motion can be obtained simply by applying Q learning method as one of reinforcement learning methods. Our previous studies show relation between the successful giant swing motion and rewards which are defined by robot postures or kinetic and potential energy. This method can obtain the successful giant swing motion when proper learning conditions are given unless statistical performance is occurred due to Non-Markov property. As this method is useful to investigate selective ability of learning knowledge because this experiment has various motions, we focused on its selective ability. This paper clarifies the selective ability in the view point that directional giant rotation can be obtained from the learning knowledge whose rotation has both a forward and a back rotation. Its selective method is proposed, whose results show significant improvement of the forward rotation given by penalty of the backward rotation. However, its improvement of the backward rotation is not so significant. Discussion for the selective ability is given by using both simulation results and experimental results.