Learning-Based Adaptive User Selection in Millimeter Wave Hybrid Beamforming Systems

Junghoon Kim, Matthew Andrews

TL;DR

This paper tackles the challenge of efficiently selecting users for transmission in millimeter wave (mmWave) hybrid beamforming systems, where limited RF chains restrict the number of simultaneous users. They propose a machine learning (ML)-based approach that learns an optimal user selection strategy, achieving a better trade-off between user fairness and computational efficiency compared to traditional methods.

What Problem Does This Paper Address?

Millimeter wave (mmWave) frequencies offer vast bandwidth for 5G and beyond, but suffer from high path loss. To overcome this, base stations (BS) use hybrid beamforming, combining analog and digital techniques with a limited number of RF chains. This limitation means the BS can only serve a subset of users at any given time. Efficiently selecting which users to serve is crucial for maximizing system performance, particularly for proportional fairness (PF), which ensures no user is starved of service. Existing methods struggle to balance the complex optimization required for user selection with the rapid channel variations and strict time constraints (a few milliseconds) in mmWave systems, often sacrificing either fairness or speed.

Methods

The authors propose a two-timescale protocol to manage the distinct timescales of mmWave channel characteristics. A long timescale is used for selecting analog beams (which are relatively stable), while a short timescale handles user selection and digital beamforming (which must adapt to faster channel changes).

For user selection, they first introduce baseline algorithms:

  • Greedy Algorithm: Iteratively adds users that maximize the PF metric, offering good performance but being computationally expensive.
  • Top-k Algorithm: A simpler heuristic that selects the top-k users based on a simplified rate metric, offering speed but lower PF.

To bridge the gap between performance and speed, they propose a Machine Learning (ML)-based user selection algorithm using deep neural networks (DNNs). The DNN takes as input the analog beam indices, real-time effective channels, and scheduling weights. It outputs a user selection vector, which is then processed to determine the final set of users. This ML approach aims to learn the complex relationships between these inputs and the optimal user set for maximizing PF, offering a faster inference time after training.

Datasets

Not explicitly discussed as distinct datasets. The authors simulate the system and generate training and testing data based on their defined simulation setup, which includes parameters like the number of users, BS antenna configuration, and channel models.

Key Insights and Ideas

  • Two-Timescale Protocol is Key: Separating analog beam selection (long timescale) from user selection and digital beamforming (short timescale) is crucial for efficient operation in mmWave systems.
  • ML for Trade-off: Machine learning, specifically DNNs, can effectively learn complex user selection strategies that strike a balance between proportional fairness and computational efficiency, outperforming traditional greedy and heuristic methods.
  • Channel Interference Mitigation: The ML-based approach, by learning from data, implicitly learns to mitigate channel interference among selected users, leading to a higher minimum chordal distance between their channel vectors.
  • Practical Constraints Matter: The paper emphasizes the need for algorithms that operate within the tight coherence time of mmWave channels, making computationally intensive methods like the pure greedy algorithm impractical for real-time deployment.

Key Results

  • The ML-based algorithm achieves a 71% accuracy in predicting the optimal user selection.
  • In terms of Proportional Fairness (PF), the ML-based algorithm performs comparably to the greedy algorithm, significantly outperforming the top-N and top-1 algorithms.
  • Crucially, the ML-based algorithm achieves this performance with a significantly lower run time than the greedy algorithm, making it suitable for real-time operation. It offers a better trade-off than the adaptive top-k algorithm.
  • The ML-selected users exhibit a high minimum chordal distance, indicating effective interference mitigation.

Limitations and Future Work

  • Accuracy vs. PF: While the ML model's direct prediction accuracy is 71%, the paper highlights that the resulting PF performance is still competitive, suggesting that direct accuracy might not be the sole metric for evaluating such systems.
  • DNN Architecture: The paper explores different DNN architectures, finding that a two-hidden-layer network with 500 and 200 neurons performs best. However, further optimization of DNN architecture and input features could potentially yield even better results.
  • Future Work: The authors suggest further improving the ML-based algorithm to significantly outperform baselines across various metrics. They also mention exploring the use of multiple data streams per user to fully utilize RF chains for enhanced performance.

Summary updated

Abstract

We consider a multi-user hybrid beamforming system, where the multiplexing gain is limited by the small number of RF chains employed at the base station (BS). To allow greater freedom for maximizing the multiplexing gain, it is better if the BS selects and serves some of the users at each scheduling instant, rather than serving all the users all the time. We adopt a two-timescale protocol that takes into account the mmWave characteristics, where at the long timescale an analog beam is chosen for each user, and at the short timescale users are selected for transmission based on the chosen analog beams. The goal of the user selection is to maximize the traditional Proportional Fair (PF) metric. However, this maximization is non-trivial due to interference between the analog beams for selected users. We first define a greedy algorithm and a "top-k" algorithm, and then propose a machine learning (ML)-based user selection algorithm to provide an efficient trade-off between the PF performance and the computation time. Throughout simulations, we analyze the performance of the ML-based algorithms under various metrics, and show that it gives an efficient trade-off in performance as compared to counterparts.
BibTeX
@inproceedings{kim2023learning,
  author = {Junghoon Kim and Matthew Andrews},
  title = {{Learning-Based Adaptive User Selection in Millimeter Wave Hybrid Beamforming Systems}},
  booktitle = {ICC},
  year = {2023},
  doi = {10.1109/icc45041.2023.10279582},
}
Figure