Deep-learning-based precoding in multiuser MIMO downlink channels with limited feedback
share
We propose a deep-learning-based channel quantization, feedback, and precoding method for downlink multiuser multiple-input multiple-output systems. In the proposed system, the traditional codebook-based channel quantization process for limited feedback is handled by a receiver deep neural network (DNN) for each user. The precoder selection process is handled by a transmitter DNN for the base station. At each receiver DNN, a binarization layer is adopted to emulate the channel quantization process and enable end-to-end learning. However, during training, receiver DNNs with the binarization layer can be trapped at a poor local minimum because of inaccurate gradients caused by the binarization layer. To address this, we consider a method of knowledge distillation, in which the existing DNNs are jointly trained with an additional auxiliary transmitter DNN. By using the auxiliary DNN as a teacher network, the receiver DNNs can additionally exploit lossless gradients, which is useful in avoiding a poor local minimum. Moreover, through joint training, the existing DNNs can be generalized including the quantization loss from binarization. All DNNs at the associated users and transmitter are trained offline in an end-to-end manner, with the aid of the auxiliary transmitter DNN. The purpose of the end-to-end learning is to determine the precoding matrices that maximize the downlink sum rate. Our DNN-based precoding scheme can achieve a significantly higher downlink rate compared to traditional linear precoding with codebook-based limited feedback, for the same number of feedback bits, particularly when the number of receive antennas is greater than one.
READ FULL TEXT