In distributed cognitive radio networks, cognitive radio devices which perform the channel sensing individually, are seriously affected by radio channel environments such as noise, shadowing and fading such that they can not property satisfy the maximum allowable interference level to the primary user. In the paper, we propose a Q-learning based channel access scheme for cognitive radios so as to satisfy the maximum allowable interference level to the primary user as well as to improve the throughput of cognitive radio by opportunistically accessing on the idle channels. In the proposed scheme, the pattern of channel usage of the primary user will be learned through Q-learning during the pre-play learning step, and then the learned channel usage pattern will be utilized for improving the sensing performance during the Q-learning normal operation step. Through the simulation, it is shown that the proposed scheme can provide bettor performance than the conventional energy detector in terms of the interference level to primary user and the throughput of cognitive radio under both AWGN and Rayleigh fading channels.