Proactive UAV Network Slicing for URLLC and Mobile Broadband Service Multiplexing

Peng Yang, Xing Xi, Kun Guo, Tony Q. S. Quek, Jingxuan Chen, Xianbin CaoIEEE JSAC 2021 · DOI

TL;DR

This paper tackles the challenge of efficiently serving diverse communication needs (ultra-reliable low-latency for drone control and high-bandwidth for mobile users) simultaneously on a single UAV network. They propose a novel "proactive" network slicing approach that leverages distributed learning for user location prediction and DNNs for channel gain estimation, combined with a Lyapunov-based optimization framework to dynamically manage resources.

What Problem Does This Paper Address?

The increasing integration of Unmanned Aerial Vehicles (UAVs) into 5G and future 6G networks presents a significant challenge: how to support a multitude of services with vastly different requirements (e.g., ultra-reliable low-latency communication (URLLC) for drone control versus mobile broadband (MBB) for ground users) on a single, resource-constrained UAV platform. Traditional approaches would require separate networks for each service, which is impractical. Network slicing offers a solution by creating virtual, isolated networks on a shared infrastructure. However, applying this to UAVs is complex due to:

  • Dynamic User Locations: Ground users are mobile, making it difficult to predict their needs and allocate resources effectively.
  • Inaccurate Channel Models: Standard channel gain models are often too simplistic for real-world urban environments where UAVs operate, leading to suboptimal performance.
  • Time-Consuming Slice Management: Creating and configuring network slices is a slow process, which is problematic for services with strict latency requirements.
  • Sequence-Dependent Decisions: The optimal resource allocation at any given time depends on past decisions and future predictions, making it a challenging sequential decision problem.

This paper aims to address these challenges by developing a proactive UAV network slicing strategy that can dynamically adapt to user mobility and service demands while ensuring both URLLC and MBB requirements are met efficiently.

Methods

The paper proposes a multi-faceted approach combining learning and optimization techniques:

  • Distributed Echo State Network (ESN) for User Location Prediction: To overcome the challenge of unknown user locations, they employ a distributed learning method. Each UAV and the base station (BS) act as agents, training local ESN models based on their partial user location data. These local models are then aggregated at the BS to form a global model, which is broadcast back to the UAVs. This distributed approach avoids the need to send large amounts of raw data to a central server, reducing network overhead.
  • Deep Neural Networks (DNNs) for Channel Gain Estimation: To obtain accurate and analytically tractable channel gain models, DNNs are used. Separate DNNs are trained for the Base Station-to-UAV (BtU) and UAV-to-Ground (UtG) links. These DNNs take relevant environmental and signal information as input and output estimated channel gain coefficients. This allows for more realistic channel modeling compared to traditional simplified models.
  • Lyapunov-Based Optimization Framework: To handle the sequential decision-making nature of the problem and ensure long-term performance guarantees, a Lyapunov-based optimization framework is employed. This framework decomposes the complex optimization problem into a series of smaller, manageable subproblems that can be solved iteratively. Virtual queues are introduced to track constraint violations, and the framework aims to minimize a drift-plus-penalty function, balancing optimality with constraint satisfaction.
  • Successive Convex Approximation (SCA) and Alternative Optimization: The subproblems generated by the Lyapunov framework often involve non-convex constraints. To address this, the paper utilizes SCA techniques to approximate non-convex functions with convex ones, allowing for efficient optimization. An alternative optimization scheme is also used to handle mixed-integer programming problems.

Datasets

  • Realistic Pedestrian Movement Dataset: The paper utilizes a dataset of 12,000 Twitter posts collected near Oxford Street, London, on March 14, 2018. This dataset provides GPS-related position information for mobile users, which is used to simulate user movement and train the distributed ESN for location prediction.
  • 3GPP Models: Standard models from the 3GPP (3rd Generation Partnership Project) are used for simulating channel gains (e.g., urban macro model for path loss) and antenna characteristics.

Key Insights and Ideas

  • Proactive Slicing: The core idea is to "proactively" slice the network based on predicted user locations and channel conditions, rather than reactively configuring slices as demands arise. This addresses the time-consuming nature of slice creation.
  • Distributed Learning for Scalability: Using a distributed ESN for user location prediction is crucial for scalability. It avoids overwhelming a central server with data from numerous UAVs and users, making the system more robust and efficient.
  • DNNs for Realistic Channel Modeling: The use of DNNs to estimate channel gains allows for a more accurate representation of the complex radio environment, which is essential for optimizing resource allocation in urban settings.
  • Lyapunov Optimization for Long-Term Guarantees: The Lyapunov framework provides a principled way to handle the dynamic and sequential nature of the problem, ensuring that performance objectives (like energy efficiency and fairness) are met over the long term, while also satisfying critical URLLC constraints.

Key Results

  • Accurate User Location Prediction: The distributed ESN method demonstrates good accuracy in predicting user trajectories, with Mean Squared Error (MSE) for predicted trajectories not exceeding 0.8. The method can re-capture user locations even after rapid changes in movement direction.
  • Effective Channel Gain Estimation: DNNs achieve good accuracy in estimating both BtU and UtG channel gains, with loss values decreasing significantly during online learning. The estimation error is manageable even with fast-changing channel conditions due to user and UAV movement.
  • Virtual Queue Stability: The Lyapunov-based optimization framework successfully stabilizes virtual queues, ensuring that URLLC and MBB service requirements are met. Queue stability values tend towards zero, indicating that constraints are satisfied.
  • Improved Energy Efficiency and Fairness: The proposed RE2FS algorithm significantly outperforms benchmark algorithms (Static UAV-based and Circular Trajectory-based) in terms of energy efficiency and Jain's fairness index. For instance, it achieves up to 14.58% higher energy efficiency and 9.13% higher fairness index compared to benchmarks.
  • Adaptability to UAV and User Density: The performance of RE2FS scales well with an increasing number of users and UAVs, demonstrating its robustness in various deployment scenarios.

Limitations and Future Work

  • User Movement Prediction Limitations: While the ESN method is effective, it can temporarily lose track of users when their movement directions change very rapidly. The paper suggests that the method can re-capture future locations after retraining on new samples.
  • 3D Movement of UAVs: The current model primarily considers 2D movement for UAVs, with fixed altitudes. Future work could explore optimizing UAV movement in 3D space.
  • Cooperation between BS and UAVs: The paper mentions that cooperation between BSs and UAVs to serve ground users is left for future research due to space limitations.
  • Generalization to Other Environments: While the urban environment is considered challenging, extending the framework to other environments (e.g., rural, maritime) could be a future direction.
  • Other Realistic Pedestrian Datasets: The paper notes that the algorithm accommodates other realistic pedestrian movement datasets, suggesting potential for further validation.

Summary updated

BibTeX
@article{yang2021proactive,
  author = {Peng Yang and Xing Xi and Kun Guo and Tony Q. S. Quek and Jingxuan Chen and Xianbin Cao},
  title = {{Proactive UAV Network Slicing for URLLC and Mobile Broadband Service Multiplexing}},
  journal = {IEEE J. Sel. Areas Commun.},
  year = {2021},
  doi = {10.1109/jsac.2021.3088624},
}
Figure