In cognitive radio networks, rendezvous algorithms play an important role. They aim to promise secondary users to meet on the same channel in a bounded time. But these works often neglect the influence of primary user activities. On the other way, there may be more collisions as the number of secondary users raised up. The aim of this paper is to design a method to let secondary users keep meeting on a better channel. We first model rendezvous problem as a game, then approximate the desire equilibriums using reinforcement learning. The convergence in some simple cases is proved. Some extra mechanisms are applied to handle the case of multiple secondary user pairs. Through simulations, we show the presented reinforcement learning technique is able to converge to channels with higher channel idle probabilities.