A competition-cooperation teacher-student model is proposed for knowledge transfer between channels for sleep staging. The teacher model was trained simultaneously with competitor and cooperator student models. One extra “competitor” student model was added to the traditional teacher-student model. The performance of the teacher model can be optimized during the competition with the competitor model. The overfitting of the student model is greatly reduced in cooperation with the teacher model. A three-step competition-cooperation training process was proposed to optimize the performance of both the teacher model and the student models. The experimental results demonstrate a 2.7% and a 1.6% improvement of the accuracy for the teacher model and the student model, respectively. A stateof-the-art sleep staging accuracy of 83.7% with single-channel input and 85.9% with multi-channel inputs was achieved by the proposed algorithm.