Wireless medium access control (MAC) protocol identification based on power detection plays an important role in cognitive radio and cognitive electronic warfare. To improve accuracy, features like collision proportion estimation and Fisher statistic are derived. Aiming at the problem of low accuracy when the target network is different from training samples, the selective ensemble method based on Q-learning is utilized. The identification system is constructed from a set of selected base classifiers to improve generation capacity. The proposed method is tested with samples of four MAC protocols by OPNET simulation. Experimental results show that the features proposed in this article increase discrimination ability of different MACs. Moreover, when the target network is different from training samples, the selective ensemble method has higher accuracy than the single classifier and the ensemble of all classifiers.