Joint User Scheduling and Beamforming Design for Multiuser MISO Downlink Systems

Shiwen He, Jun Yuan, Zhenyu An, Wei Huang, Yongming Huang, Yaoxue Zhang

TL;DR

This paper tackles the complex challenge of simultaneously deciding which users to serve (scheduling) and how to shape their signals (beamforming) in a multiuser wireless system. By moving beyond traditional separate optimizations, the authors introduce a Graph Neural Network (GNN) approach that achieves near-optimal performance with significantly higher computational efficiency and better adaptability to changing network conditions.

What Problem Does This Paper Address?

In modern wireless networks, especially those supporting the Internet of Things (IoT), base stations must manage limited power and bandwidth to serve many devices. This involves two core tasks: User Scheduling (US) (deciding which users get a connection) and Beamforming (BF) (adjusting the phase and amplitude of signals at multiple antennas to minimize interference).

Traditionally, these are treated as separate problems because solving them jointly is "NP-hard"—the discrete choice of users coupled with the continuous variables of beamforming creates a mathematical nightmare. Furthermore, most existing research relies on Shannon’s capacity theory, which assumes infinitely long data packets. This paper focuses on Ultra-Reliable Low-Latency Communication (URLLC), where packets are short (finite blocklength), making the math even more non-convex and difficult to solve using standard tools.

Methods

The authors propose two distinct solutions to the joint US-BF problem:

  1. SCA-USBF (The Optimization Approach): This is a model-driven algorithm based on Successive Convex Approximation. It uses "uplink-downlink duality" to transform the downlink problem into a more manageable virtual uplink problem. It approximates the discrete "scheduled/not scheduled" decision using continuous variables and iteratively solves convex sub-problems. While accurate, it is computationally expensive.
  2. J-USBF (The Learning Approach): This is the paper's primary innovation. It uses a GNN-based model called JEEPON (Joint User Scheduling and Power Allocation Network).
  3. Graph Representation: The network is modeled as a graph where users are nodes and the interference between them are edges.
  4. Unsupervised Learning: Instead of needing "optimal labels" (which are hard to calculate), the network is trained using a Primal-Dual Learning Framework (PDLF). It treats the constraints (like power limits and minimum rates) as penalties in a Lagrangian relaxation, allowing the model to learn the physics of the environment directly.
  5. Analytical Integration: The GNN predicts scheduling and power, while the beamforming vectors are calculated using an analytical MMSE (Minimum Mean Square Error) solution, combining the best of deep learning and classical signal processing.

Datasets

The study relies on simulated link-level data rather than public datasets, which is standard for this sub-field of wireless communications.

  • Environment: A single base station with 32 antennas serving 30 to 50 single-antenna users.
  • Channel Model: Users are randomly distributed in a circular cell (50m to 200m radius) with Rayleigh small-scale fading and standard path-loss models.
  • Training/Testing: They generated 5,000 training samples and 500 testing samples to evaluate convergence and generalizability.

Key Insights and Ideas

  • Topology Matters: Standard Deep Neural Networks (DNNs) often struggle as network size grows. By using a GNN, the authors exploit the inherent "graph" nature of wireless interference, allowing the model to generalize to different numbers of users without retraining from scratch.
  • Duality as a Bridge: The use of uplink-downlink duality isn't just a math trick; it allows the GNN to work in a "virtual" space where the constraints are easier to satisfy during the learning process.
  • Efficiency via Iteration: To avoid the high cost of matrix inversions required for SINR (Signal-to-Interference-plus-Noise Ratio) calculations, the authors implemented a matrix iteration lemma, reducing computational complexity from $O(N^3)$ to $O(N^2)$.

Key Results

  • Performance: The J-USBF (GNN) approach achieves performance very close to the SCA-USBF optimization (often within 2-5% of the greedy baseline) but operates orders of magnitude faster during the "online" phase.
  • Generalizability: A model trained on a specific user distribution (e.g., users 100m-120m away) still performed effectively when tested on different distributions (e.g., 80m-100m), proving it learned the underlying spatial logic rather than just memorizing a map.
  • Reliability: In testing, the GNN-based approach satisfied the complex URLLC rate constraints in over 97.7% of cases, a high mark for an unsupervised learning model.

Limitations and Future Work

  • Perfect CSI Assumption: The paper assumes the base station has "Perfect Channel State Information" (it knows exactly how the signal will bounce and fade for every user). In reality, CSI is often imperfect or delayed. The authors acknowledge that moving toward statistical CSI is a necessary next step.
  • Single-Cell Focus: The current model only considers one base station. Real-world deployments involve multi-cell interference, which would significantly complicate the graph representation and the beamforming design.
  • Signaling Overhead: While the GNN is fast, the "cost" of acquiring the channel data to feed into the GNN remains high. Future work could look at reducing this signaling overhead.

Summary updated

Abstract

In multiuser communication systems, user scheduling and beamforming (US-BF) design are two fundamental problems that are usually studied separately in the existing literature. In this work, we focus on the joint US-BF design with the goal of maximizing the set cardinality of scheduled users, which is computationally challenging due to the non-convex objective function and the coupled constraints with discrete-continuous variables. To tackle these difficulties, a successive convex approximation based US-BF (SCA-USBF) optimization algorithm is firstly proposed. Then, inspired by wireless intelligent communication, a graph neural network based joint US-BF (J-USBF) learning algorithm is developed by combining the joint US and power allocation network model with the BF analytical solution. The effectiveness of SCA-USBF and J-USBF is verified by various numerical results, the latter achieves close performance and higher computational efficiency. Furthermore, the proposed J-USBF also enjoys the generalizability in dynamic wireless network scenarios.
BibTeX
@article{he2023joint,
  author = {Shiwen He and Jun Yuan and Zhenyu An and Wei Huang and Yongming Huang and Yaoxue Zhang},
  title = {{Joint User Scheduling and Beamforming Design for Multiuser MISO Downlink Systems}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2023},
  url = {https://openreview.net/forum?id=LCrmkKvgpa},
}

Loading related content...

Figure