Learning of Prediction

Summary Our humans can know a motion of a target, and obtain appropriate actions based on prediction, even if both target and ourselves mvoe together. Acquisition of such actions, which needs some context information, is important for a robot in dynamic environment. In case of target and robot move together, the robot must compensate its motion because the sensory signals are influenced by the robot motion. If the knowledge about the change of the visual signals according to the robot motion is not given in advance, it is difficult for a robot to know the target movement. Since the prediction is done to realize more appropriate actions towards some purpose, it can be regarded as "a way to achive a certain purpose". Then, it is considered that such actions can be obtained according to the necessity by learning of the action for the purpose.

Here, a capturing task of a moving object is employed as an environment, in which both robot and target move. The appropriate actions based on prediction in this task are learned by the combination of Elman-type recurrent neural network and reinforcement learning. Moreover, the effect of the sensory motion is focused on. Three kinds of sensory motions, 1) looking to a constant direction in absolute coordinates, 2)keeping the object in the center, and 3) fixed on the robot, are employed, and the learning results are compared. Simulation result is shown that the robot can obtain appropriate actions based on prediction faster in the case of 1) and 2) than 3).

Reference
5. 前原伸一, 杉坂政典, 柴田克成:
移動物体の捕獲行動学習におけるセンサ動作の必要性,
電子情報通信学会技術報告, NC2001-153, pp. 159-166, 2002.3
(in Japanese)

4. Shin'ich Maehara, Masanori Sugisaka and Katsunari Shibata:
Effectiveness of Sensory Motion in the Learning of Capturing Task of a Moving Object
Proc. of AROB (Int'l Sympo. on Artificial Life and Robotics) 7th, pp. 46-49, 2002.1

3. 前原伸一, 杉坂政典, 柴田克成:
移動物体の捕獲行動学習におけるセンサ動作の必要性,
計測自動制御学会情報・システム部門学術講演会2001, pp. 13-18, 2001. 11.
(in Japanese)

2. 西岡忠相，柴田克成，伊藤宏司：
モデル型ＴＤ強化学習による動的環境での行動獲得,
第11回自律分散システムシンポジウム資料, pp. 285-288, 1999.1
(in Japanese)

1. 西岡忠相，柴田克成，伊藤宏司：
動的環境における移動ロボットの行動獲得,
第16回日本ロボット学会学術講演会予稿集, Vol. 1, pp. 331-332，1998.9
(in Japanese)

Return to my home page (English)
Return to my home page (Japanese)