| 摘要 |
PROBLEM TO BE SOLVED: To provide a learning control system which has high predicting capability and is capable of fast learning and adaptive to even real-time control over a complicated domain. SOLUTION: Provided is a behavior learning controller capable of learning input/output relationship according to its behavior. This controller includes a sensor which acquires information on an external environment, a sense evaluation portion which discriminates whether the state of a controlled system is stable from sensor information and outputs a reinforcement signal corresponding to a discrimination result, a sensor information state separation portion which carries out reinforced learning based upon the sensor information and the reinforcement signal supplied from the sense evaluation portion, weights the sensor information with higher sense evaluation more, and classifies the sensor information into a low-dimensional state, and a behavior learning portion which learns the relationship between the state supplied from the sensor information state separation portion and a corresponding behavior control command and outputs the behavior control command. Thus, the reinforced learning is introduced in the sensor information separation portion, so that effective state classifications can be expected to be learnt in a short time. COPYRIGHT: (C)2006,JPO&NCIPI
|
