MACC'97 Session: RoboCup
Applying reinforcement learning to a high-level and cooperative action problem causes the following problems: 1) An agent needs to follow its best policy for acquiring a cooperative action in spite of the fact that the agent must reinforce various policies to leave its local optimal policy. 2) Practicable policies of high-level actions tend to be very few because a high-level action is much restricted compared with a low-level action. In order to solve these problems, this paper proposes ``observational reinforcement'' method in which an agent guesses the optimal policy from observation and reinforces it. This method enables an agent to reinforce various policies while it follows its best policy, and to reinforce a hopeful policy which is impracticable currently. We ascertained the effectiveness of this method from experiments in the RoboCup agents' position problem.
reinforcement learning, multiagent, neural networks, RoboCup
PS
file(+gzip) (in Japanese)