[Paper Review] Deep Learning-Empowered Predictive Beamforming for IRS-Assisted Multi-User Communications

Chang Liu, Xuemeng Liu+4 more|arXiv (Cornell University)|Apr 26, 2021

Advanced Wireless Communication Technologies20 references4 citations

TL;DR

This paper proposes a deep learning-based predictive beamforming scheme for IRS-assisted multi-user communications that reduces training overhead by predicting IRS phase shifts using historical line-of-sight (LoS) channel features. By combining a location-aware convolutional long short-term memory network (LA-CLNet) for predictive phase shift estimation and an ICSI-aware fully-connected neural network (IA-FNN) for instantaneous transmit beamforming, the method achieves sum-rate performance close to the genie-aided scheme with full CSI, significantly reducing channel estimation overhead.

ABSTRACT

The realization of practical intelligent reflecting surface (IRS)-assisted multi-user communication (IRS-MUC) systems critically depends on the proper beamforming design exploiting accurate channel state information (CSI). However, channel estimation (CE) in IRS-MUC systems requires a significantly large training overhead due to the numerous reflection elements involved in IRS. In this paper, we adopt a deep learning approach to implicitly learn the historical channel features and directly predict the IRS phase shifts for the next time slot to maximize the average achievable sum-rate of an IRS-MUC system taking into account the user mobility. By doing this, only a low-dimension multiple-input single-output (MISO) CE is needed for transmit beamforming design, thus significantly reducing the CE overhead. To this end, a location-aware convolutional long short-term memory network (LA-CLNet) is first developed to facilitate predictive beamforming at IRS, where the convolutional and recurrent units are jointly adopted to exploit both the spatial and temporal features of channels simultaneously. Given the predictive IRS phase shift beamforming, an instantaneous CSI (ICSI)-aware fully-connected neural network (IA-FNN) is then proposed to optimize the transmit beamforming matrix at the access point. Simulation results demonstrate that the sum-rate performance achieved by the proposed method approaches that of the genie-aided scheme with the full perfect ICSI.

Motivation & Objective

• To address the high training overhead in IRS-assisted multi-user communication (IRS-MUC) systems due to large-scale IRS reflection elements.
• To reduce reliance on frequent channel state information (CSI) estimation by leveraging historical channel features for predictive beamforming.
• To design a low-overhead beamforming framework that maintains high spectral efficiency despite user mobility and imperfect CSI.
• To develop a deep learning architecture that jointly exploits spatial and temporal channel correlations for accurate IRS phase shift prediction.
• To achieve sum-rate performance close to the theoretical upper bound with minimal instantaneous CSI feedback.

Proposed method

• A location-aware convolutional long short-term memory network (LA-CLNet) is proposed to predict IRS phase shifts by modeling spatial and temporal features of historical LoS channels.
• The LA-CLNet uses convolutional layers to extract spatial correlation across IRS elements and LSTM layers to capture temporal dynamics from past channel states.
• Only a low-dimensional multiple-input single-output (MISO) channel estimation is required at the access point (AP), based on the predicted IRS phase shifts.
• An ICSI-aware fully-connected neural network (IA-FNN) is designed to optimize the transmit beamforming matrix at the AP using the effective instantaneous CSI.
• The IA-FNN is trained to map the effective CSI (after predictive IRS beamforming) to optimal precoding vectors, maximizing sum-rate.
• The end-to-end framework operates in an unsupervised manner, learning from historical LoS data without requiring real-time CSI feedback during inference.

Experimental results

Research questions

• RQ1 Can predictive beamforming based on historical LoS channel data reduce the training overhead in IRS-MUC systems?
• RQ2 How effectively can a deep learning model exploit both spatial and temporal channel features to predict IRS phase shifts?
• RQ3 To what extent can predictive IRS phase shifts enable high sum-rate performance with minimal instantaneous CSI feedback?
• RQ4 How does the proposed method compare to baseline schemes relying on outdated or random phase shifts?
• RQ5 Can the proposed method achieve performance close to the genie-aided scheme with perfect instantaneous CSI?

Key findings

• The proposed DLPB method achieves a sum-rate performance within 5 dB of the genie-aided scheme with full instantaneous CSI, demonstrating near-optimal performance.
• The sum-rate of the proposed method increases with the Rician factor β and approaches the upper bound at β = 10 dB, indicating improved beam alignment with stronger LoS components.
• At β = 8 dB and P = 30 dBm, the proposed method achieves a 5 dB sum-rate gain over the naive DL and random phase-shift schemes.
• The method outperforms the naive DL and random PS schemes significantly, especially in high-mobility scenarios, due to effective use of historical channel data.
• The sum-rate of the genie-aided DL-ICSI method decreases with increasing β due to reduced channel degrees of freedom, while the proposed method improves under high LoS dominance.
• The framework reduces CSI feedback overhead by replacing full-scale IRS channel estimation with a low-dimensional MISO estimation after predictive phase shift design.