Learning-Based Fronthaul Compression for Uplink Cloud Radio Access Networks

Abstract

This paper investigates the uplink signal dimension reduction problem for a user-centric cloud radio access network, in which each single-antenna user communicates with the central processor (CP) through a cluster of remote radio heads (RRHs). To reduce the fronthaul traffic, each RRH applies a compression matrix to reduce the dimension of the received signal before relaying it to the CP. However, the optimal design of the compression matrices requires significant communication overhead for transmitting the high-dimensional channel state information (CSI) matrices from the RRHs to the CP. To address this issue, this paper proposes a deep learning framework to first learn a sub-optimal compression matrix at each RRH based on the local CSI, then iteratively refine the learned compression matrix using a meta-learning-based gradient method. To reduce the communication cost for CSI sharing and gradients transmission, this paper proposes an efficient signaling scheme that only requires the transmission of low-dimensional effective CSI and its gradient between the CP and each RRH. Furthermore, a meta-learning-based gated recurrent unit (GRU) network is proposed to reduce the number of signaling transmission rounds. For the sum-rate maximization problem, simulation results show that the proposed two-stage neural network can perform closely to the fully cooperative global CSI-based benchmark with significantly reduced communication overhead. Moreover, using the first stage alone can already outperform the existing local CSI-based benchmark.