Preservation and Utilization of Reinforcement Learning Robot's Strategies Using Probablistic Networks
Toshiyuki Yasuda, Kazuhiro Ohkura · TRANSACTIONS OF THE JAPAN SOCIETY OF MECHANICAL ENGINEERS Series C · 2007
Reinforcement learning (RL) has been proven to be a promising approach to behavior acquisition of an autonomous robot. However, it holds the unwanted character of over-fitting such that a RL robot often loses its stable behavior to a change occurred in an environment after successful learning. To overcome this problem, we propose the mechanism for a robot that transforms an implicit knowledge collected as a result of RL into the form of a probabilistic network model as an explicit knowledge and preserves it for future use. We expects that a robot tends to show more robust behavior against various changes in an environment with the increase of the knowledges. The computer simulations are conducted to illustrate the effectiveness of the proposed knowledge collecting mechanism. The physical experiments with a small autonomous mobile robot called Khepera are also examined to validate the effectiveness of the proposed method.