Abstract
The development in information technology, especially computational performance, enables us to have the intelligent systems such as humanoid robot and entertainment robot in real world. These robots have to adapt their behavior depending on the dynamically changing environment. Some kinds of learning capabilities are implemented in such systems. Reinforcement learning (RL) is one of the most actively investigated algorithms, and Q-learning is one of the most popular ones. Q-learning is generally needs quite long time for learning behaviors, so many kinds of improvements for effective learning have been proposed.
Considering that these robots live with us in our daily life, it is also important to act without making us keep waiting, since we are not so patient. Reducing computational cost while executing some behaviors is important to achieve these properties. Recent studies in brain science suggest that memory based predictions may be the keys to realize saving the activities in the brain. In this paper, we try to discuss its architecture to realize these characteristics, based on reinforcement learning and associative memories.
