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.
Introduction
In modern wireless networks, particularly those supporting the Internet of Things (IoT) and Ultra-Reliable Low-Latency Communication (URLLC), base stations must efficiently manage limited power and bandwidth to serve a high density of devices. This necessitates the simultaneous optimization of two core tasks: User Scheduling (US), which determines which subset of users to connect, and Beamforming (BF), which shapes the spatial signals to maximize throughput while minimizing interference.
Traditionally, these problems are solved sequentially or independently because their joint optimization is NP-hard. The coupling of discrete scheduling decisions with continuous beamforming variables, combined with the non-convex nature of finite blocklength (short packet) regimes typical of URLLC, creates a significant mathematical challenge that standard optimization tools struggle to solve efficiently.
Methodology
This research proposes two distinct frameworks to address the joint US-BF problem: a model-driven optimization approach and a data-driven learning approach.
1. SCA-USBF: The Optimization Approach
The Successive Convex Approximation (SCA-USBF) algorithm serves as a high-accuracy benchmark. It utilizes the principle of uplink-downlink duality to transform the complex downlink problem into a more tractable virtual uplink scenario. By approximating the discrete scheduling indicators with continuous variables and iteratively solving convex sub-problems, the algorithm navigates the non-convex objective space. While effective, the iterative nature of SCA makes it computationally intensive, often proving too slow for real-time applications in dynamic environments.
2. J-USBF: The Learning Approach
To overcome the latency of traditional optimization, the authors developed J-USBF, a learning-based framework utilizing a Graph Neural Network (GNN) architecture termed JEEPON (Joint User Scheduling and Power Allocation Network).
- Graph Representation: Unlike standard deep neural networks, the GNN models the wireless network as a graph where users are nodes and inter-user interference represents edges. This allows the model to exploit the spatial topology of the network.
- Unsupervised Learning: The model is trained using a Primal-Dual Learning Framework (PDLF). Instead of requiring expensive "optimal" labels for training, it treats constraints (such as power limits and minimum rate requirements) as penalties within a Lagrangian relaxation. This allows the network to learn the underlying physics of the wireless environment directly from the channel data.
- Analytical Integration: To ensure both speed and accuracy, the GNN predicts scheduling and power allocation, while the beamforming vectors are derived using an analytical Minimum Mean Square Error (MMSE) solution. This hybrid approach combines the flexibility of deep learning with the theoretical rigor of classical signal processing.
Experimental Setup
The algorithms were evaluated using simulated link-level data. The environment consisted of a single base station equipped with 32 antennas serving between 30 and 50 single-antenna users. Users were randomly distributed within a circular cell (50m to 200m radius) under Rayleigh small-scale fading and standard path-loss models. The training phase utilized 5,000 samples, while 500 samples were reserved for testing convergence and generalizability.
Key Insights and Results
Performance and Efficiency
The J-USBF approach demonstrated performance remarkably close to the SCA-USBF optimization, typically staying within 2-5% of the greedy baseline. Crucially, J-USBF operates orders of magnitude faster during the online inference phase, making it suitable for real-time deployment. To further enhance efficiency, the authors implemented a matrix iteration lemma that reduced the computational complexity of SINR calculations from $O(N^3)$ to $O(N^2)$.
Generalizability and Reliability
A significant advantage of the GNN-based J-USBF is its ability to generalize. A model trained on a specific user distribution (e.g., 100m-120m distance) maintained high performance when tested on different spatial distributions (e.g., 80m-100m). This indicates that the GNN learned the fundamental spatial logic of interference rather than merely memorizing specific coordinates. Furthermore, the model satisfied complex URLLC rate constraints in over 97.7% of test cases, a high reliability mark for an unsupervised learning model.
Conclusion and Future Work
This work demonstrates that joint user scheduling and beamforming can be effectively managed by combining GNNs with analytical signal processing. While the proposed J-USBF framework shows great promise, the authors identify several areas for future research:
- Imperfect CSI: Moving from the assumption of perfect Channel State Information (CSI) to statistical or delayed CSI to better reflect real-world conditions.
- Multi-Cell Coordination: Expanding the model to account for multi-cell interference, which would increase the complexity of the graph representation.
- Signaling Overhead: Investigating methods to reduce the overhead required to acquire the channel data necessary for GNN input.
Summary updated