Hybrid Precoding for Beamspace MIMO Systems With Sub-Connected Switches: A Machine Learning Approach

Ting Ding, Yongjun Zhao, Lixin Li, Dexiu Hu, Lei ZhangIEEE Access 2019 · DOI

Introduction

Millimeter-wave (mmWave) massive MIMO is a cornerstone of next-generation wireless communications, promising significant increases in data throughput. However, the practical implementation of these systems is hindered by the "RF chain bottleneck." Ideally, each antenna would be paired with a dedicated Radio Frequency (RF) chain, but the resulting cost and power consumption for hundreds of antennas are prohibitive.

To mitigate this, hybrid precoding architectures are employed, combining high-dimensional analog beamforming with low-dimensional digital precoding. Traditionally, the analog component relies on networks of Phase Shifters (PSs), which, while effective for beam steering, are significant energy consumers.

Traditional hybrid precoding architecture for mmWave massive MIMO.

Proposed System Architecture

This research proposes a more energy-efficient alternative: replacing power-hungry phase shifters with a sub-connected switch selecting network integrated with a lens antenna array. In this "sub-connected" model, each RF chain is linked only to a specific subset of antennas, reducing hardware complexity and internal power loss.

Proposed sub-connected switch selecting network with a lens antenna array and 1-bit phase shifter detail.

The use of switches introduces a binary constraint (on/off), transforming the analog beamforming design into a complex combinatorial optimization problem. Traditional gradient-based methods are ill-suited for such discrete search spaces.

The Hybrid Cross-Entropy (HCE) Algorithm

To optimize the switch configurations, the authors developed a Hybrid Cross-Entropy (HCE) algorithm, drawing from machine learning optimization techniques. The process follows an iterative refinement cycle:

  1. Probabilistic Modeling: The state of each switch is modeled as a Bernoulli random variable, representing the probability of the switch being active.
  2. Sampling and Evaluation: Multiple candidate configurations are sampled based on these probabilities. For each sample, the optimal digital precoder is determined using a classical Wiener Filter (WF) to maximize the sum-rate.
  3. Iterative Refinement: The algorithm identifies "elitist" candidates—those yielding the highest sum-rates—and updates the underlying probability distribution.
  4. Smoothing for Global Convergence: To prevent premature convergence to local minima, the HCE algorithm incorporates a smoothing constant ($\alpha$). This ensures a more robust exploration of the solution space compared to standard Cross-Entropy methods.

Performance Evaluation

The proposed HCE-based scheme was evaluated using the Saleh-Valenzuela channel model, simulating a base station with 256 antennas and 16 RF chains serving 16 users.

Key findings include:

  • Energy Efficiency (EE): The switch-based architecture demonstrated a dramatic improvement in Energy Efficiency (bits per Joule) compared to traditional PS-based systems. By eliminating the high power draw of phase shifters, the system achieves a much higher throughput-to-power ratio.
  • Sum-Rate Performance: Despite the reduced degrees of freedom inherent in switches, the HCE algorithm achieved a higher sum-rate than conventional antenna selection (AS) and standard CE methods.
  • Computational Efficiency: The algorithm demonstrated rapid convergence, typically reaching an optimized state within approximately 20 iterations.

Conclusion and Future Outlook

While switch-based networks may not match the absolute peak sum-rate of high-resolution phase-shifter arrays due to coarser beamforming granularity, the gains in energy efficiency make them a compelling choice for green 5G/6G infrastructure. Future research will explore the integration of deep learning to further optimize digital signal processing and evaluate the performance of these architectures in highly dynamic, real-world environments.

Summary updated

BibTeX
@article{ding2019hybrid,
  author = {Ting Ding and Yongjun Zhao and Lixin Li and Dexiu Hu and Lei Zhang},
  title = {{Hybrid Precoding for Beamspace MIMO Systems With Sub-Connected Switches: A Machine Learning Approach}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2944061},
}
Figure