Learn to Allocate Resources in Vehicular Networks

Liang Wang, Hao Ye, Le Liang, Geoffrey Ye Li

Abstract

Resource allocation has a direct and profound impact on the performance of vehicle-to-everything (V2X) networks. Considering the dynamic nature of vehicular environments, it is appealing to devise a decentralized strategy to perform effective resource sharing. In this paper, we exploit deep learning to promote coordination among multiple vehicles and propose a hybrid architecture consisting of centralized decision making and distributed resource sharing to maximize the long-term sum rate of all vehicles. To reduce the network signaling overhead, each vehicle uses a deep neural network to compress its own observed information that is thereafter fed back to the centralized decision-making unit, which employs a deep Q-network to allocate resources and then sends the decision results to all vehicles. We further adopt a quantization layer for each vehicle that learns to quantize the continuous feedback. Extensive simulation results demonstrate that the proposed hybrid architecture can achieve near-optimal performance. Meanwhile, there exists an optimal number of continuous feedback and binary feedback, respectively. Besides, this architecture is robust to different feedback intervals, input noise, and feedback noise.

Introduction

In modern vehicular networks, spectrum is an increasingly scarce resource. Vehicle-to-Vehicle (V2V) links must often share frequency bands with Vehicle-to-Infrastructure (V2I) links, creating a complex resource allocation challenge. Traditional centralized schemes, while high-performing, suffer from a "signaling storm" due to the massive amount of Channel State Information (CSI) required by the base station (BS). Conversely, fully distributed schemes minimize overhead but often lead to interference and suboptimal global performance. This research proposes a middle ground: a hybrid architecture that achieves near-optimal performance with minimal signaling overhead by leveraging deep learning for intelligent information compression.

Proposed Architecture

The core of the proposed system is a BS-Aided Spectrum Sharing Architecture that utilizes two distinct types of neural networks working in tandem to balance local observation with global coordination.

Hybrid architecture for BS-aided spectrum sharing.

The architecture consists of two primary components:

  1. Distributed Compression (Vehicle Side): Each V2V pair employs a Deep Neural Network (DNN) to process local environmental data, such as channel gains and interference levels. Instead of transmitting raw data, the DNN extracts and compresses the most relevant features into a compact feedback vector.
  2. Centralized Decision-Making (BS Side): The BS aggregates these compressed vectors and inputs them into a Deep Q-Network (DQN). This Reinforcement Learning (RL) agent determines the optimal channel allocation for all vehicles to maximize the network's long-term sum rate.

To further enhance real-world applicability, a quantization layer is integrated into the vehicle-side DNNs. This allows the system to learn how to convert continuous feedback into a limited number of bits (binary feedback) without significant performance loss. The entire pipeline is trained end-to-end, enabling the vehicles and the BS to develop a "common language" for efficient resource management.

Experimental Evaluation

The architecture was evaluated using a V2X simulator built according to 3GPP standards (TR 36.885), simulating an urban environment with 4 V2V pairs and 4 V2I links.

Performance comparison of DNN-RL vs. optimal and random schemes.

The results demonstrate that the proposed DNN-RL scheme achieves approximately 95% to 99% of the performance of a computationally intensive "brute-force" optimal search. As shown in the evaluation, the system vastly outperforms random channel selection, which typically achieves less than 40% of the optimal sum rate.

Key findings include:

  • Minimal Overhead: Peak performance is reached with just 3 real-valued feedback variables or 36 bits of binary feedback per vehicle.
  • Robustness: The system maintains high performance despite feedback delays or significant noise (up to 30dB noise-to-feedback ratio), showcasing its resilience in dynamic vehicular environments.

Conclusion

This paper presents a novel hybrid deep learning framework for V2X resource allocation. By shifting from raw data transmission to intelligent feature feedback, the proposed architecture effectively mitigates signaling overhead while maintaining near-optimal throughput. Future work will explore scaling the system to larger vehicle densities using Multi-Agent Reinforcement Learning (MARL) and incorporating strict Quality-of-Service (QoS) constraints for safety-critical applications.

Summary updated

Abstract

Resource allocation has a direct and profound impact on the performance of vehicle-to-everything (V2X) networks. Considering the dynamic nature of vehicular environments, it is appealing to devise a decentralized strategy to perform effective resource sharing. In this paper, we exploit deep learning to promote coordination among multiple vehicles and propose a hybrid architecture consisting of centralized decision making and distributed resource sharing to maximize the long-term sum rate of all vehicles. To reduce the network signaling overhead, each vehicle uses a deep neural network to compress its own observed information that is thereafter fed back to the centralized decision-making unit, which employs a deep Q-network to allocate resources and then sends the decision results to all vehicles. We further adopt a quantization layer for each vehicle that learns to quantize the continuous feedback. Extensive simulation results demonstrate that the proposed hybrid architecture can achieve near-optimal performance. Meanwhile, there exists an optimal number of continuous feedback and binary feedback, respectively. Besides, this architecture is robust to different feedback intervals, input noise, and feedback noise.
BibTeX
@misc{wang2019learn,
  author = {Liang Wang and Hao Ye and Le Liang and Geoffrey Ye Li},
  title = {{Learn to Allocate Resources in Vehicular Networks}},
  year = {2019},
  eprint = {1908.03447},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1908.03447},
}
Figure