Abstract
Modern wireless networks, particularly those supporting 5G and 6G applications like ultra-reliable low-latency communication (uRLLC), face significant challenges in real-time resource allocation. Traditional iterative optimization methods, while mathematically rigorous, are often too computationally intensive for millisecond-level decision-making. This paper presents UPGDNet, an unsupervised deep unrolling framework that combines the structural advantages of Projection Gradient Descent (PGD) with the efficiency of Graph Neural Networks (GNNs). By unrolling iterative steps into neural network layers and employing a novel constraint-splitting strategy, UPGDNet achieves near-optimal performance with a 99% reduction in computational overhead compared to traditional benchmarks.
Introduction
Resource allocation problems in wireless networks, such as beamforming and power control, are typically formulated as non-convex constrained optimization problems. These problems are characterized by "coupled" variables, where the configuration of one user impacts the performance of others. While traditional algorithms like Successive Convex Approximation (SCA) provide reliable solutions, their iterative nature leads to high latency, making them unsuitable for real-time uRLLC applications like autonomous driving or remote surgery.
Deep learning offers a faster alternative, but standard "black-box" Deep Neural Networks (DNNs) often struggle to strictly satisfy the complex physical constraints inherent in wireless systems. UPGDNet addresses this by "unrolling" the PGD algorithm, creating a model-driven AI architecture that inherits mathematical interpretability while gaining the speed of neural inference.
The UPGDNet Framework
Deep Unrolling and Architecture
UPGDNet transforms the iterations of a PGD algorithm into the layers of a deep network. Each layer represents a single iteration, where the network learns to optimize the transition from one state to the next.
As shown in the architecture above, the framework introduces several key innovations:
- Constraint Splitting: Constraints are divided into two categories. Category 1 ($C_1$) includes simple, convex, and decoupled constraints handled by a standard mathematical projection ($\Pi_{C_1}$). Category 2 ($C_2$) involves complex, non-convex, or coupled constraints. Instead of direct projection, the network learns a "perturbation vector" via $\mathcal{N}_x$ to nudge the solution toward the feasible region.
- Learnable Step-Sizes: Rather than using a fixed gradient step-size ($\eta$), the network employs a sub-module ($\mathcal{N}_\eta$) to learn the optimal step-size for every specific network realization.
- Permutation Equivariance: To handle varying network topologies, UPGDNet utilizes Heterogeneous Weighted Graph Convolutional Networks (HWGCNs). This ensures that the model treats users as nodes in a graph, maintaining consistent performance regardless of user ordering.
Unsupervised Primal-Dual Learning
A significant advantage of UPGDNet is that it does not require "labels" or pre-solved optimal solutions for training. It utilizes an unsupervised primal-dual learning approach, minimizing a Lagrangian loss function. This function penalizes both poor objective performance (e.g., low data rates) and violations of the $C_2$ constraints, allowing the network to discover optimal strategies autonomously.
Experimental Results
The framework was evaluated using USRMNet, a specific implementation designed for downlink multiuser uRLLC systems. Simulations were conducted across varying user counts ($K=4$ to $10$) and signal-to-noise ratios (SNR) using Rayleigh fading channel models.
- Performance: USRMNet consistently achieved over 99% of the Weighted Sum Rate (WSR) produced by the high-performing traditional baseline, HeBF.
- Efficiency: The computational savings were dramatic. In a 10-user scenario, the traditional iterative method required approximately $7.63 \times 10^7$ operations. In contrast, USRMNet required only $6.93 \times 10^5$ operations—representing a 99% reduction in computational complexity.
- Generalization: The model demonstrated robust generalization, maintaining high performance even when tested on user spatial distributions that were not present during the training phase.
Conclusion
UPGDNet represents a significant step forward in bridging the gap between mathematical optimization and deep learning for wireless communications. By unrolling the PGD algorithm and intelligently splitting constraints, the framework provides a high-speed, near-optimal solution for complex resource allocation. Future work will focus on extending this architecture to high-mobility scenarios and further optimizing graph filter banks for extremely dense networks with hundreds of devices.
Summary updated