Predictive Relay Selection: A Cooperative Diversity Scheme Using Deep Learning

Wei Jiang, Hans D. SchottenarXiv 2021 · DOI · arXiv

Abstract

In this paper, we propose a novel cooperative multi-relay transmission scheme for mobile terminals to exploit spatial diversity. By improving the timeliness of measured channel state information (CSI) through deep learning (DL)-based channel prediction, the proposed scheme remarkably lowers the probability of wrong relay selection arising from outdated CSI in fast time-varying channels. It inherits the simplicity of opportunistic relaying by selecting a single relay, avoiding the complexity of multi-relay coordination and synchronization. Numerical results reveal that it can achieve full diversity gain in slow-fading channels and substantially outperforms the existing schemes in fast-fading wireless environments. Moreover, the computational complexity brought by the DL predictor is negligible compared to off-the-shelf computing hardware.

Introduction

In cooperative wireless communication, multiple relay nodes assist a source in transmitting data to a destination, effectively creating a virtual multi-antenna system that enhances reliability through spatial diversity. Among various strategies, Opportunistic Relay Selection (ORS) is favored for its simplicity, as it selects the single best relay based on current Channel State Information (CSI).

However, the rise of high-mobility applications—such as drones, high-speed trains, and Vehicle-to-Everything (V2X) systems—presents a significant challenge: fast-fading channels. In these environments, the channel conditions change so rapidly that by the time a relay is selected and begins its transmission, the CSI used for that selection is already "outdated." This leads to incorrect relay choices and a catastrophic loss of diversity gain. To address this, we introduce Predictive Relay Selection (PRS), a scheme that leverages deep learning to anticipate future channel states.

The PRS Framework and Architecture

The PRS scheme integrates a Deep Learning (DL) predictor into a dual-hop decode-and-forward (DF) relay system. Unlike traditional methods that react to past measurements, each relay in the PRS framework uses a local neural network to predict the future channel magnitude ($|h|$).

Architecture of the Deep Learning CSI Predictor integrated within the relay node.

The system utilizes a timer-based selection protocol. During the signaling phase, relays with the highest predicted future channel gain have their timers expire first, allowing them to "claim" the transmission slot for the subsequent frame. This ensures that the relay with the best actual channel at the time of transmission is the one being used.

Deep Learning for Channel Prediction

The core of the PRS scheme is the DL CSI Predictor. While simple Recurrent Neural Networks (RNNs) often struggle with long-term dependencies in time-series data, we found that Long Short-Term Memory (LSTM) networks are ideally suited for this task.

Through extensive testing, a 2-layer LSTM architecture with 25 neurons per layer was identified as the "Goldilocks" zone—providing high prediction accuracy without the over-fitting risks associated with deeper models. A key insight of this work is that for relay selection, predicting the complex phase of the channel is unnecessary; predicting the magnitude (or SNR) is sufficient. This simplification allows the neural network to operate with real-valued weights, significantly reducing computational overhead.

Performance Evaluation

We evaluated the PRS scheme using a simulated dataset based on the Jakes model, representing a fast-fading environment with a maximal Doppler frequency of 100 Hz. The results demonstrate that PRS effectively restores the diversity gain that is lost by traditional ORS in high-mobility scenarios.

Outage probability comparison between PRS, ORS, and OSTC schemes under different delay conditions.

As shown in the performance analysis, PRS maintains near-optimal performance even as the signal-to-noise ratio (SNR) increases. At an outage probability of $10^{-3}$ with a 3ms feedback delay, PRS outperforms the Opportunistic Space-Time Coding (OSTC) scheme by approximately 8dB. Furthermore, at an SNR of 20dB, PRS achieves a spectral efficiency of 3.5 bps/Hz, a 35% improvement over the 2.6 bps/Hz achieved by traditional ORS.

Crucially, the "computational tax" of this AI integration is minimal. The LSTM model requires only approximately 15.3 MFLOPS. On a standard digital signal processor (DSP) like the TI C6678, this task consumes less than 0.01% of the available processing power, proving that PRS is highly viable for real-world hardware implementation.

Conclusion and Future Work

Predictive Relay Selection offers a robust solution to the outdated CSI problem in fast-fading wireless channels. By shifting from reactive to predictive selection, PRS maintains high reliability and capacity in high-mobility environments without the synchronization complexity of multi-relay schemes. Future research will explore the robustness of these predictors in non-idealized channel models and investigate how this framework can be integrated into the emerging "AI-native" air interfaces of 6G networks.

Summary updated

Abstract

In this paper, we propose a novel cooperative multi-relay transmission scheme for mobile terminals to exploit spatial diversity. By improving the timeliness of measured channel state information (CSI) through deep learning (DL)-based channel prediction, the proposed scheme remarkably lowers the probability of wrong relay selection arising from outdated CSI in fast time-varying channels. It inherits the simplicity of opportunistic relaying by selecting a single relay, avoiding the complexity of multi-relay coordination and synchronization. Numerical results reveal that it can achieve full diversity gain in slow-fading channels and substantially outperforms the existing schemes in fast-fading wireless environments. Moreover, the computational complexity brought by the DL predictor is negligible compared to off-the-shelf computing hardware.
BibTeX
@misc{jiang2021predictive,
  author = {Wei Jiang and Hans D. Schotten},
  title = {{Predictive Relay Selection: A Cooperative Diversity Scheme Using Deep Learning}},
  year = {2021},
  eprint = {2102.03325},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2102.03325},
}
Figure