2004 年 22 巻 5 号 p. 672-689
Reinforcement learning is very interesting for robot learning. However, there are some significant problems in applying conventional reinforcement learning algorithms to the robot with many degrees of freedom, because the size of exploration space increase exponentially with increase of degrees of freedom, and it makes it impossible to accomplish learning process. On the other hand, animals and humans can learn and accomplish various tasks using many redundant degrees of freedom of the body in spite of the exploration space is very huge. In this paper, we consider how to solve the state explosion problem in applying the reinforcement learning to the redundant robot and propose new framework of reinforcement learning, which is inspired by the body image of animals, by summarizing our previous works of reinforcement learning. To demonstrate the effectiveness of proposed method, simulations and experiments have been carried out and as a result effective behaviors have been obtained.