Profit sharing reducing the occurrences of accidents by predicted action-safety degree

This paper aims to reduce the number of occurrences of accidents during reinforcement learning(RL). RL is a powerful tool for autonomous acquisition of behavior without knowledge of dynamics of environment. However, because conventional algorithms require huge number of experiences of taking action to learn its goodness, the robots can not avoid having accidents during learning process. Such accidents may causes a breakdown and make the robots impossible to accomplish a task. Therefore, learning algorithm requires reducing the occurrence of accidents. This paper focuses on knowledge of robot designers to reduce the occurrences of accidents during the learning by using sensor information relating to accidents. This paper proposes a new PS-based algorithm which considers predicted action-safety degree to select, in order to reduce the occurrences of accidents during the learning. As a result, the effectiveness of proposed method was confirmed by simulation experiments.