Deep CSI Compression and Coordinated Precoding for Multicell Downlink Systems

Abstract

This work proposes a deep-learning (DL) based coordinated precoder design for multicell downlink systems with rate-limited exchange of channel state information (CSI) among base-stations (BSs). Two CSI compression techniques are proposed, one based on a binarized convolutional neural network (CNN) and one based on a learned vector-quantization (VQ) codebook. The former utilizes a CNN-based CSI feature extractor to directly compute the binary feature vector that is to be exchanged with other BSs. The latter utilizes a DL-based VQ codebook to encode the CSI feature vector that is obtained at the output of the feature extractor. In both cases, each BS takes the rate-limited CSI received from other BSs as input to a precoder network that produces the normalized precoding vectors and the transmit powers using a multitask learning architecture. By using solutions of the weighted minimum mean square error (WMMSE) algorithm as the output labels, end-to-end training of both the CSI compression and transmit precoder networks is performed jointly at all BSs. By doing so, the CSI compression networks will be able to extract the CSI features that are most effective for precoder computation at the BSs. Our simulation results show that the proposed schemes can achieve weighted sum rates close to that in the full CSI scenario, even when the number of exchanged bits is small, and outperform existing random VQ methods.