Abstract
As 5G and Internet of Things (IoT) networks expand, Device-to-Device (D2D) communication becomes essential for reducing cellular load. However, traditional centralized power allocation methods introduce significant signaling overhead and latency. This paper presents DPADIC, a distributed deep learning framework that enables D2D transmitters to autonomously determine transmit power using only local location data. By utilizing a shared Deep Neural Network (DNN) trained on global objectives but executed locally, the system achieves near-optimal throughput with minimal latency.
Introduction
In modern cellular architectures, D2D communication allows devices to share spectrum with the primary network (underlaying). The primary challenge is managing interference between D2D pairs and the base station (eNB). Centralized control by the eNB is the standard approach, but it scales poorly with the explosion of IoT devices. The resulting "control burden" leads to high latency, which is incompatible with 5G's ultra-reliable low-latency communication (URLLC) requirements.
Methodology: The DPADIC Framework
The proposed DPADIC (Distributed Power Allocation using DNN with Interference to eNB Constraint) shifts the computational load from the center to the edge.
Distributed Architecture
The framework operates in two distinct phases:
- Centralized Training: During the offline phase, a central entity uses global information (locations of all devices) to train a single universal Deep Neural Network. The goal is to maximize the aggregate throughput of the entire cell while respecting interference constraints.
- Distributed Execution: Once deployed, every D2D device carries an identical copy of the pre-trained model. Each device inputs its own local (x, y) coordinates and independently calculates its transmit power across multiple subchannels.
Neural Network Design
The core of DPADIC is a Multi-Layer Perceptron (MLP) featuring seven layers. The input consists of the (x, y) coordinates for both the transmitter and receiver of a D2D pair. Instead of using traditional labeled datasets, the model employs the mathematical throughput objective function as its loss function. This "label-free" approach uses ReLU-based penalty terms to enforce power limits and interference thresholds for the eNB.
Key Insights
- Location as a Proxy for Channel State: The DNN learns to map simple geometric coordinates to the underlying physics of path loss and interference, effectively using location as a proxy for complex channel gain measurements.
- Scale Invariance and Consistency: By using a single universal model, the system ensures consistent behavior across devices. Furthermore, models trained on a specific number of D2D pairs (e.g., 8) demonstrate robust performance even when the actual number of active devices fluctuates.
Results and Performance
The researchers evaluated DPADIC using a hexagonal cell simulator with 5 million random device placements. Key findings include:
- Efficiency: DPADIC achieved spectral efficiency and throughput comparable to the iterative IADRMPIC algorithm, a known benchmark for distributed rate maximization.
- Latency: Power allocation is reduced to a single forward pass of the DNN, drastically shortening decision times compared to iterative or centralized methods.
- Interference Control: The model successfully learned to suppress power for devices near the cell edge or the eNB, maintaining interference levels below critical thresholds (e.g., -140 dBW).
Conclusion
DPADIC demonstrates that deep learning can effectively decentralize complex optimization problems in wireless networks. By replacing iterative signaling with local inference, the framework meets the low-latency demands of next-generation IoT. Future work will focus on incorporating small-scale fading effects and adapting the model for high-mobility scenarios.
Summary updated