An improved Q-learning algorithm using experience sharing for multi-robot system

This paper proposes an improved Q-learning algorithm using experience sharing to improve the learning efficiency of traditional Q-learning. Traditional Q-learning as a classic reinforcement learning (RL) has simple operation and small size of state-action space, and can be applied in multi-robot system (MRS). But compared with multiagent reinforcement learning algorithm, traditional Q-learning lacks information exchange with other agents. Experience sharing which imitates human thinking is a good way for solving this problem. By experience sharing each robot can share with other robots'Q values through a gradual learning process using ε-greedy policy to get learning experience with probability 1-ε. Robot soccer is adopted as test platform and simulation result shows that the improved Q-learning algorithm outperforms the traditional Q-learning algorithm. ©, 2015, Binary Information Press. All right reserved.