The University of Edinburgh -
Division of Informatics
Forrest Hill & 80 South Bridge

MSc Thesis #9827

Title:Reinforcement Learning in Bayesian Networks using Stochastic Gradient Ascent
Date: 1998
Abstract:This dissertation discussed the possibility of the Reinforcement Learning in Bayesian Networks using the Stochastic Gradient Ascent introduced in [Yamamura & Onozuka 98]. After explaining the mechanism of Reinforcement Learning and Bayesian Networks, three experiments to confirm the behaviour of the Stochastic Gradient Method on a Bayesian Network were conducted. These results implied that more complicated Bayesian Networks are not easy to learn. some of the methods to overcome complex Bayesian Networks were suggested. In addition to that, this methodology was applied to a simulated robot in the Khepera robot simulator and the merit of using prior knowledge in the robot controller was discussed. In the task of moving around the environment as much as possible, the robot with prior knowledge made poorer performance that that without prior knowledge at the end of the experiment. The possible cause includes the structure of the Bayesian Network used. To discover the cause, the extension of the software to handle more complicated Bayesian Network is necessary.

[Search These Pages] [DAI Home Page] [Comment]