| 103C3 |
|---|
| Verification of Body Growth Effect on Reinforcement Learning in a Simple Standing-up Task |
| AUTHORS |
| Daiki KIYOSUKE, Masanori SUGISAKA, Katsunari SHIBATA (Oita University) |
| ABSTRACT |
| Why does human grow up? As one solution the authors believe that the body growth accelerates the learning of actions, such as standing up in higher forms of lives. While, the slow learning is a serious problem in Reinforcement Learning because of the trial and error factor. In this paper, it was verified whether the body growth accelerates the learning in a simple standing-up problem. It was shown that even though the pendulum could not learn to standing-up when its arm was long, it could learn when the arm became longer from short state during the learning. The learning is very sensitive to the parameters, especially the scaling of the joint torque as the growth of the arm length. In this paper, the torque was scaled to be proportional to the fourth power of the arm length by the consideration of the torque to keep the pendulum laterally against the gravity. If the torque was scaled to be proportional to the fifth power, the pendulum with the long arm could learn to stand up. However, from the consideration of the validity of scaling, it is thought that the proportion to the fifth power is too large. At the last it was shown that in the case of balancing Barto”Ēs indication is suitable but in the case of standing-up body growth is effective. |
Back to Program
Time Table
Top of This Site
Annual Conference 2002