DeepMux: Deep-Learning-Based Channel Sounding and Resource Allocation for IEEE 802.11ax
Pedram Kheirkhah Sangdeh, Huacheng Zeng
Abstract
MU-MIMO and OFDMA are two key techniques in IEEE 802.11ax standard. Although these two techniques have been intensively studied in cellular networks, their joint optimization in Wi-Fi networks has been rarely explored as OFDMA was introduced to Wi-Fi networks for the first time in 802.11ax. The marriage of these two techniques in Wi-Fi networks creates both opportunities and challenges in the practical design of MAC-layer protocols and algorithms to optimize airtime overhead, spectral efficiency, and computational complexity. In this paper, we present DeepMux, a deep-learning-based MU-MIMO-OFDMA transmission scheme for 802.11ax networks. DeepMux mainly comprises two components: deep-learning-based channel sounding (DLCS) and deep-learning-based resource allocation (DLRA), both of which reside in access points (APs) and impose no computational/communication burden on Wi-Fi clients. DLCS reduces the airtime overhead of 802.11 protocols by leveraging the deep neural networks (DNNs). It uses uplink channels to train the DNNs for downlink channels, making the training process easy to implement. DLRA employs a DNN to solve the mixed-integer resource allocation problem, enabling an AP to obtain a near-optimal solution in polynomial time. We have built a wireless testbed to examine the performance of DeepMux in real-world environments. Our experimental results show that DeepMux reduces the sounding overhead by 62.0% ~ 90.5% and increases the network throughput by 26.3% ~ 43.6%.