The invention discloses a network selection method based on a Q-learning algorithm. The network selection method includes the steps of (1) initializing a Q value table and setting a discount factor gamma and a learning rate alpha; (2) determining a business type k and load rates of two current networks when a set time is up so as to obtain a current state sn; (3) selecting a useable action from anaction set A and recording such action as well as next network state sn+1; (4) computing an immediate return function r according to the network state after the selected action is implemented; (5), updating a Q value function Qn (s, a), and gradually decreasing the learning rate alpha to 0 according to rules of an inverse proportional function; (6), repeating the steps (2)-(5) until Q values areconverged, in other words, a difference value of the Q values before and after updating is smaller than a threshold value; (7) returning to the step (3) to select the action and accessing to an optimal network. The network selection method based on the Q-learning algorithm is capable of decreasing a voice business block rate and a data service packet loss rate and increasing average network throughout.