Abstract: How to achieve efficient routing in the dynamical and complex network is one of current research hotspots. Q-learning, a frequently used reinforcement learning method, which can solve the optimal control problem in unknown environment by continuously interacting with the environment, is able to achieve on-line learning task. A rank-based Q-routing algorithm (RQ routing) is proposed. RQ routing algorithm, taking Q-learning algorithm as learning framework, and preserving the efficiency of the Q-routing algorithm, introduces the rank function that can be dynamically calculated to represent the priority of the current state in the scene, so as to solve the optimal solution of the route selection, which can avoid long waiting queue, reduce network congestion and improve the transmission speed. The rank function in the RQ routing algorithm is flexible. People can use different rank functions to meet the needs of various scenes, ensure the better generalization ability of the algorithm, and overcome the inflexibility of the traditional Q-routing application scene. The experiment verifies the effectiveness of the algorithm.

Key words: reinforcement learning, Q-learning, Q-routing, QoS routing, computer network