 |
    |
On-Line Model-Based Continuous State Reinforcement Learning Using Background Knowledge
Bernhard Hengst
University of New South Wales, Computer Science and Engineering, Australiabernhardh@cse.unsw.edu.au
http://www.cse.unsw.edu.au
Abstract. Without a model the application of reinforcement learning to control a dynamic system can be hampered by several shortcomings. The number of trials needed to learn a good policy can be costly and time consuming for robotic applications where data is gathered in real-time. In this paper we describe a variable resolution model-based reinforcement learning approach that distributes sample points in the state-space in proportion to the effect of actions. In this way the base learner economises on storage to approximate an effective model. Our approach is conducive to including background knowledge to speed up learning. We show how different types of background knowledge can used to speed up learning in this setting. In particular, we show good performance for a weak type of background knowledge by initially overgeneralising local experience.
Keywords: Artificial Intelligence, Machine Learning, Reinforcement Learning, Robotics, Background Knowledge, Function Approximation, Continuous State, Model-Based
LNAI 7691, p. 851 ff.
lncs@springer.com
© Springer-Verlag Berlin Heidelberg 2012