More effective reinforcement learning by introducing sensory information

Abstract

Among various reinforcement learning methods, Q-learning is particularly useful for mobile robots, because its value function is a function of a state and an action. The state here represents location and orientation of a mobile robot. We propose to introduce sensory signals into reinforcement learning to increase its learning speed and the probability of reaching a goal, and to decrease the probability of collision. A key idea is to directly reduce a value function at other states than the current state of a mobile robot based on sensory signals. Computer simulation demonstrates that the number of goals reached increases more than 2 times faster both in a simple environment and in a complex environment than that by conventional Q-learning.