Reduction of state space on reinforcement learning by sensor selection

In recent years, there are many researches about applying reinforcement learning to robot. A problem of reinforcement learning is learning time. In reinforcement learning, information from sensors is projected to state space. A robot learns correspondence each state in state space and each action and finds the best correspondence. When state space is expanded depending on the number of sensors, correspondence which a robot should learn increases. That is why, it takes time to learn the best correspondence. In this paper, we focus on importance of sensors for facing task. Important sensors for achieving task are different on each task. It is not indispensable for a robot to use all installed sensors for facing task. It is hopeful that state space consists of only important sensors for facing task. Using state space which consists of only important sensors, a robot can learn faster than the case of using all installed sensors. Therefore, we will propose a faster learning system which a robot can autonomously select important sensors for facing task and constructs state space for only important sensors. We define a measure of importance of sensor for facing task. The measure is coefficient of correlation between sensor value of each sensor and reward on reinforcement learning. A robot decides important sensors based on correlation. A robot reduce state space based on important sensors. A robot can learn efficiently by reduced state space. We confirm effectiveness of a system which we propose on a simulation.