Abstract
Efficient radio resource allocation is critical for meeting the low-latency and high-throughput requirements of 5G New Radio (NR). This paper proposes a novel solution using a 2-Dimensional Hopfield Neural Network (2D-HNN) to solve the resource block (RB) assignment problem. By reformulating the Generalized Proportional Fair (GPF) scheduling objective as a neural network energy function, the system achieves near-optimal allocation through parallelized convergence. Simulations show that this approach maintains high user fairness (median >92%) and provides a tunable trade-off between fairness and total cell throughput.
Introduction
In 5G NR, the Resource Block (RB) serves as the fundamental unit of time and frequency allocation. Unlike previous standards, 5G NR introduces significant flexibility through multiple "numerologies," allowing subcarrier spacings to vary from 15 kHz to 480 kHz. This flexibility complicates the task of the Base Station (BS), which must decide RB assignments for multiple User Equipments (UEs) every millisecond.
The core challenge lies in balancing total network throughput with user fairness. Prioritizing users with the strongest signals maximizes throughput but leaves cell-edge users with poor service. Conversely, strict fairness can degrade overall network efficiency. Finding an optimal solution in real-time is computationally intensive, necessitating new approaches beyond traditional sequential optimization.
Methodology: The 2D-Hopfield Neural Network Approach
The researchers address this by integrating Generalized Proportional Fair (GPF) scheduling with a 2-Dimensional Hopfield Neural Network (2D-HNN).
The Mapping Trick
The central innovation is the mapping of the scheduling problem onto the HNN's architecture. A Hopfield network is a recurrent neural network that evolves its state to minimize an internal "energy function." By mathematically aligning the GPF objective function with this energy function, the "stable state" reached by the network corresponds to the optimal resource allocation.
As shown in the figure above, the allocation is represented as a matrix where each neuron corresponds to a potential assignment of a specific RB to a specific UE. Because HNN neurons can update simultaneously, this approach is inherently parallel, making it significantly faster than traditional iterative algorithms and suitable for the millisecond-level requirements of 5G.
Tunable Fairness
The GPF formula includes a parameter, $\alpha$, which acts as a "control knob" for network operators. A higher $\alpha$ (e.g., 1.0) prioritizes fairness, ensuring that even users with poor channel conditions receive resources. A lower $\alpha$ (e.g., 0.2) shifts the priority toward maximizing the total data rate of the cell.
Simulation Environment
The proposed model was validated using simulated data, a standard practice in radio resource management research. The simulation featured a single Base Station with UEs distributed randomly. Key parameters included:
- Resource Availability: 100 RBs per time slot.
- Channel Modeling: User data rates were calculated using the Shannon capacity model, accounting for noise and varying channel gains across different 5G numerologies (15 kHz and 60 kHz).
Results and Discussion
The 2D-HNN demonstrated a robust ability to converge on fair and efficient solutions.
Fairness Performance
With $\alpha = 1$, the system achieved a median fairness value of 92%. The distribution of fairness across various simulation runs shows that most scenarios resulted in fairness between 87% and 95%, indicating highly consistent performance for cell-edge users.
Throughput vs. Fairness Trade-off
The simulation confirmed the effectiveness of the $\alpha$ parameter:
- High Throughput Mode ($\alpha = 0.2$): The total sum data rate for eMBB users ranged between 20 Mbps and 24 Mbps.
- High Fairness Mode ($\alpha = 1$): While the total data rate slightly decreased (ranging from 15 Mbps to 22 Mbps), the service quality for the "worst-off" users improved significantly.
Conclusion
By treating resource allocation as a physical system seeking its lowest energy state, the 2D-HNN provides an elegant and high-speed solution for 5G NR scheduling. While the complexity of neuron interconnects remains a consideration for massive-scale hardware implementation, the inherent parallelism of the HNN offers a promising path for real-time, "online" resource management. Future research will focus on high-mobility scenarios and direct latency comparisons with existing industry heuristics.
Summary updated