Title: Modular reinforcement learning for control problems with multi sensors

Authors: Hayato Nakama; Satoshi Yamada

Addresses: Department of Intelligent Mechanical Engineering, Okayama University of Science, 1-1 Ridai-cho, Kita-ku, Okayama 700-0005, Japan. ' Department of Intelligent Mechanical Engineering, Okayama University of Science, 1-1 Ridai-cho, Kita-ku, Okayama 700-0005, Japan

Abstract: The modular reinforcement learning system, which is composed of some control modules and a selection module, was developed to apply to the task where several types of sensor information were necessary for the control. In this study, the modular reinforcement learning was applied to the task where the second order correlation of two different types of sensors must be discriminated. The target (goal) is a bar with the correct mark with a lamp, and other objects have one of them or another mark. To discriminate between the target and other objects, the |AND| condition of the light sensor and camera information must be distinguished. Since the learning efficiency was low, the iterative learning and the initial learning were proposed. As a result, the appropriate module selections and action selections were trained by the modular reinforcement learning.

Keywords: autonomous robots; robot control; learning control; modular reinforcement learning; robot learning; multiple sensors.

DOI: 10.1504/IJAMECHS.2011.043373

International Journal of Advanced Mechatronic Systems, 2011 Vol.3 No.4, pp.251 - 258

Published online: 18 Mar 2015 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article