Greedy exploration policy of Q-learning based on state balance

Q-learning is one of the successfully established algorithms for the reinforcement learning, which has been widely used to the intelligent control system, such as the control of robot pose. However, curse of dimensionality and difficulty in convergence exist in Q-learning arising from random exploration policy. In this paper, we propose a greedy exploration policy of Q-learning with rule guidance. This exploration policy can reduce the non-optimal action exploration as more as possible, and speed up the convergence of Q-learning. Simulation results indicate the effectiveness of the proposed method.