強化学習を用いた能動知覚方策の学習

抄録

We propose a learning algorithm for an autonomous robot to acquire the observation skill that is advantageous for achieving its task. We consider the robot has movement and observation agents separately; the observation agent learns a policy for providing the observation for the movement agent, which learns how to achieve tasks better. Each policy is trained separately, and the observation policy is updated by using the differential value function before and after the movement policy is learned by the observation given by itself. Experiments on 2D navigation tasks in simulation show that our algorithm is more successful than conventional methods for the situation in which agent’s view is narrow.