Learning attentive-depth switching while interacting with an agent

This paper addresses a learning system design for a robot based on an extended attention process. We consider that typical attention that consists of the position/area of a sight can be extended from the viewpoint of reinforcement learning (RL) systems. We propose an RL system that is based on extended attention. The proposed system learns to switch its attention depth according to the situations around the robot. We conducted two experiments to validate the proposed system: a capture task and a navigation task. In the capture task, the proposed system learned faster than traditional systems using switching. Q-value analysis confirmed that attention depth switching was developed in the proposed system. In the navigation task, the proposed system demonstrated faster learning in a more realistic environment. This attention switching provides faster learning for a wider class of RL systems.