LB-SciFi: Online Learning-Based Channel Feedback for MU-MIMO in Wireless LANs

Abstract

Multi-user MIMO (MU-MIMO) is a key technology for current and next-generation wireless local area networks (WLANs). While it has widely been deployed in WLANs, its potential is not fully exploited in real-world systems. This can be attributed to the large airtime overhead induced by channel acquisition in existing MU-MIMO protocols, which significantly compromises the throughput gain of MU-MIMO. In this paper, we present LB-SciFi, a learning-based channel feedback framework for MU-MIMO in WLANs. LB-SciFi takes advantage of recent advances in deep neural network autoencoder (DNN-AE) to compress channel state information (CSI) in 802.11 protocols, thereby conserving airtime and improving spectral efficiency. The key component of LB-SciFi is an online DNN-AE training scheme, which allows an AP to train DNN-AEs by leveraging the side information of existing 802.11 protocols. With this training scheme, DNN-AEs are capable of significantly lowering the airtime overhead for MU-MIMO while preserving its backward compatibility with incumbent Wi-Fi client devices. We have implemented LB-SciFi on a wireless testbed and evaluated its performance in indoor wireless environments. Experimental results show that LB-SciFi offers an average of 73% airtime overhead reduction and increases network throughput by 69% on average when compared to 802.11 feedback protocols.