A Machine Learning Approach to Achieving Energy Efficiency in Relay-Assisted LTE-A Downlink System

Hammad Hassan, Irfan Ahmed, Rizwan Ahmad, Hédi Khammari, Ghulam M. Bhatti, Waqas Ahmed, Muhammad Mahtab Alam2019 · DOI

TL;DR

This paper tackles the challenge of improving energy efficiency (EE) in LTE-A cellular networks by proposing novel machine learning-based resource block (RB) and power allocation schemes. They introduce a two-step neural network approach that optimizes EE and throughput, especially in scenarios with multiple relays, outperforming existing methods.

What Problem Does This Paper Address?

Cellular networks, particularly Long Term Evolution-Advanced (LTE-A), are significant energy consumers. While increasing data rates and user throughput is crucial, doing so without a proportional increase in energy consumption is a major design challenge. This is especially true for cell-edge users who often require more energy to achieve a decent signal quality due to path loss and interference. Furthermore, the complexity of resource allocation in advanced networks with features like relays (especially multiple relays) adds computational overhead, making real-time optimization difficult. This paper aims to bridge this gap by developing energy-efficient resource allocation strategies that are computationally feasible and effective, even in complex multi-relay scenarios.

Methods

The authors propose a multi-faceted approach that combines traditional resource allocation techniques with machine learning.

  • Single Relay Scenario:
    • They introduce the SAMM (alternating Maximum Throughput and Proportional Fairness) scheme for RB allocation.
  • They employ the Bisection-based Optimal Power Allocation (BOPA) algorithm for power allocation, which iteratively finds optimal power levels to maximize EE.
  • They compare these schemes with basic Maximum Throughput (MT) and Proportional Fairness (PF), with both equal power and BOPA.
  • Multiple Relay Scenario:
    • Clustering: A k-means clustering approach is used to determine optimal relay placement and user association, addressing the complexity introduced by multiple relays.
  • Machine Learning for Allocation: A two-step neural network (NN) approach is proposed:
  • RB Allocation: A multiclass classification neural network (trained using SAMM principles) is used to assign resource blocks.
  • Power Allocation: A two-layer feedforward neural network (trained using BOPA principles) is used to determine optimal power allocation. This NN approach is designed to reduce computational complexity compared to running the full optimization at each transmission time interval (TTI).

Datasets

The paper does not explicitly mention using publicly available datasets. Instead, the training data for the neural networks was generated from the implementation of their proposed algorithms (SAMM and BOPA) and simulations. For the RB allocation NN, 25,000 K-dimensional samples of received SNR were generated. For the power allocation NN, the training dataset consisted of received SNR and the corresponding allocated transmit power.

Key Insights and Ideas

  • The Trade-off is Real: The paper clearly demonstrates the inherent trade-off between maximizing throughput and ensuring fairness, and how energy efficiency is influenced by both.
  • Machine Learning for Complexity Reduction: The core innovation lies in using machine learning (specifically neural networks) to learn complex allocation strategies (RB and power) from simulation data. This allows for near-optimal performance without the prohibitive computational cost of traditional optimization methods in real-time.
  • Clustering for Scalability: The k-means clustering approach is a practical step towards managing the complexity of multi-relay deployments by intelligently grouping users and determining relay positions.
  • SAMM NN's Promise: The SAMM NN approach shows significant promise in achieving a good balance between energy efficiency and throughput fairness, especially in multi-relay scenarios, outperforming simpler schemes.
  • BOPA's Effectiveness: The BOPA algorithm, when combined with SAMM, consistently shows strong performance in single-relay scenarios for EE and fairness.

Key Results

  • Single Relay:
    • SAMM with BOPA achieved a 26% power saving compared to MT with BOPA.
  • The proposed SAMM scheme with equal power allocation showed a 77% increase in EE compared to MT with equal power.
  • Multiple Relays (SAMM NN):
    • The SAMM NN approach achieved a 2.07 times increase in EE compared to SAMM Equal Power, at the cost of a 0.72 times decrease in throughput.
  • In terms of fairness, SAMM NN showed improved performance, with even farthest users (9 and 10) achieving higher throughput.
  • The paper also compares their approach to LERPA [11], showing that SAMM NN offers better EE and throughput fairness.
  • Table 3 summarizes simulation results, showing SAMM NN achieving the highest EE (1.0630 Mbps/Watts) among the SAMM variants, while MT BOPA achieved the highest average throughput (4.5775 Mbps). LERPA [11] showed the best fairness index (0.5797).

Limitations and Future Work

  • Hypothetical MT: The paper notes that the MT scheme, while showing high throughput, is hypothetical because it only allocates resources to users with the highest SNR, which is not practical.
  • Interference in Multi-Relay: The paper acknowledges that interference between relays in a multi-relay scenario can negatively impact user throughput and EE, although their SAMM NN approach mitigates this to some extent.
  • Dataset Generation: The training data for the NNs was generated through simulations. The performance might vary in real-world deployments with different channel conditions and interference patterns.
  • Future Work: The paper suggests that further research could explore more advanced machine learning techniques and investigate the impact of different relay placement strategies and user association algorithms on overall system performance. The authors also mention that the complexity of the proposed algorithms is still a consideration, and further optimization might be needed for extremely large-scale networks.

Summary updated

BibTeX
@misc{hassan2019machine,
  author = {Hammad Hassan and Irfan Ahmed and Rizwan Ahmad and Hédi Khammari and Ghulam M. Bhatti and Waqas Ahmed and Muhammad Mahtab Alam},
  title = {{A Machine Learning Approach to Achieving Energy Efficiency in Relay-Assisted LTE-A Downlink System}},
  year = {2019},
  doi = {10.3390/s19163461},
}
Figure