AI-Based Channel Prediction in D2D Links: An Empirical Validation

Nidhi Simmons, Samuel Borges Ferreira Gomes, Michel Daoud Yacoub, Osvaldo Simeone, Simon L. Cotton, David E. SimmonsIEEE Access 2022 · DOI

TL;DR

This paper investigates the effectiveness of various deep learning models (LSTM, GRU, CNN, FFN) for predicting received signal strength in Device-to-Device (D2D) communication links. The study empirically validates that recurrent neural networks like LSTMs and GRUs significantly outperform other models, especially when considering the channel's coherence time, and identifies optimal input data lengths for accurate predictions.

What Problem Does This Paper Address?

Device-to-Device (D2D) communication is a key enabler for future wireless networks (5G, 6G and beyond), promising enhanced system performance and new low-latency services. A critical aspect for efficient D2D operation is the ability to predict the wireless channel's behavior, specifically the received signal strength (RSS). However, D2D links operate in complex, dynamic environments with rich scattering and mobility, making channel prediction challenging. Traditional methods often struggle with the temporal dependencies inherent in these channels. This paper aims to empirically validate and compare the performance of various AI-based deep learning techniques for D2D channel prediction, focusing on how prediction accuracy is affected by the channel's coherence time and the amount of input data used.

Methods

The authors explore four deep learning models:

  • Long Short-Term Memory (LSTM) networks: Well-suited for sequential data, LSTMs use a memory cell and gating mechanisms (input, forget, output) to capture long-term dependencies.
  • Gated Recurrent Units (GRUs): A simplified version of LSTMs, GRUs also use gating mechanisms (reset, update) to manage information flow and are effective for sequential data.
  • Convolutional Neural Networks (CNNs): Typically used for image data, 1D CNNs are applied here to extract features from time-series data by using convolutional filters.
  • Feedforward Networks (FFNs) / Dense Networks: These are standard multi-layer perceptrons where each neuron is connected to every neuron in the next layer.

A linear regression model is used as a baseline for comparison. The data is preprocessed using min-max normalization and then fed into these models using a sliding window approach to frame the time-series prediction problem as a supervised learning task. The models are trained and evaluated on real-world D2D channel measurements.

Datasets

The study utilizes real-world D2D channel measurements collected at 5.8 GHz. These measurements cover four distinct environments and scenarios:

  • Indoor open-office area (LOS and NLOS)
  • Outdoor open-space area (LOS and NLOS)

The data captures received signal strength (RSS) variations. The authors do not introduce a new benchmark dataset but rather use their collected data for empirical validation.

Key Insights and Ideas

  • Recurrent Networks Reign Supreme: LSTMs and GRUs consistently outperform FFNs, CNNs, and linear regression. This highlights their superior ability to capture the temporal dynamics and long-term dependencies crucial for D2D channel prediction.
  • Coherence Time is Key: The study emphasizes the relationship between the channel's coherence time and the required input data length. Predictions beyond the coherence time are less effective, as past channel states become uncorrelated.
  • Short Input Lengths Can Suffice: Surprisingly, in many cases, a relatively short input length (around 25 ms) achieves performance comparable to much longer input sequences (up to 100 ms). This suggests that focusing on data within the coherence time is more beneficial than simply increasing the input history.
  • GRU vs. LSTM Equivalence: For this specific task, fine-tuned GRUs and LSTMs exhibit very similar prediction performance, suggesting that the added complexity of LSTM's memory cell might not always be necessary for achieving optimal results in D2D channel prediction.
  • Human Body as a Scatterer: The paper implicitly acknowledges the complexity introduced by human proximity and movement, which are inherent to D2D scenarios and contribute to the rich scattering environments.

Key Results

  • Performance Ranking: GRUs and LSTMs consistently achieved the best prediction performance across all tested environments and scenarios, significantly outperforming FFNs, 1DCNNs, and linear regression.
  • Optimal Input Lengths: For predicting 17 ms into the future in indoor LOS, an input length of 25 ms was recommended. For outdoor LOS, predicting 23 ms into the future suggested an input length of 25 ms. For NLOS scenarios, input lengths of 25 ms and 75 ms were recommended for predicting 17-18 ms into the future.
  • Diminishing Returns: Increasing input lengths beyond approximately 25 ms often showed diminishing returns, with large input lengths (e.g., 100 ms) not always yielding better results, reinforcing the importance of coherence time.
  • Single Layer Sufficiency: For LSTMs and GRUs, a single layer with 25 hidden units generally provided optimal performance, with additional layers not significantly improving results but increasing training time.

Limitations and Future Work

  • Parameter Optimization: The paper notes that determining optimal parameters theoretically is an open research question, and parameters were determined empirically.
  • Generalizability: While the study covers various environments, the findings are based on specific measurement conditions. Future work could explore a wider range of mobility patterns, environmental conditions, and D2D configurations.
  • Computational Cost: Although LSTMs and GRUs perform well, they are computationally expensive. Future research could focus on model compression or more efficient architectures for real-time deployment.
  • Further Exploration of Input Lengths: The paper suggests that further investigation into the relationship between input length and coherence time could lead to more refined recommendations for practical D2D systems.

Summary updated

BibTeX
@article{simmons2022ai,
  author = {Nidhi Simmons and Samuel Borges Ferreira Gomes and Michel Daoud Yacoub and Osvaldo Simeone and Simon L. Cotton and David E. Simmons},
  title = {{AI-Based Channel Prediction in D2D Links: An Empirical Validation}},
  journal = {IEEE Access},
  year = {2022},
  doi = {10.1109/access.2022.3182713},
}
Figure