Reinforcement learning for the automatic control of the acrobot

Japanese - English

Keywords

Acrobot, Reinforcement learning.

Abstract

An acrobot is an underactuated robot that consists of two links and two joints, and is a simple model of horizontal bar. The first joint (wrist joint) is a free joint, and the torque can be only applied to the second joint (waist joint). Because this robot has non-linear and unstable dynamics, it is difficult to balance this robot at upright position. After 16 learning episodes (trials), a good control rule for balancing of an acrobot can be obtained by our proposed method.

Images

Horizontal bar

Control for acrobot

Acrobot

References

Yoshimoto, J., Ishii, S. & Sato, M. (2000). On-line EM reinforcement learning. in Proceedings of the IEEE-INNS-ENNS International Joint Conference on Neural Networks (IJCNN2000), pp.163-168, IEEE.
Yoshimoto, J., Ishii, S. & Sato, M. (2001). Application of reinforcement learning based on on-line EM algorithm to balancing of acrobot. Systems and Computers in Japan, 32(5), pp.12-20.