Rate-Splitting for Intelligent Reflecting Surface-Aided Multiuser VR Streaming

Rui Huang, Vincent W. S. Wong, Robert SchoberIEEE JSAC 2023 · DOI

TL;DR

This paper tackles the challenge of delivering high-quality, immersive VR experiences over wireless networks by proposing an Intelligent Reflecting Surface (IRS)-aided Rate-Splitting (RS) system. They introduce a novel deep reinforcement learning (DRL) approach, Deep-GRAIL, to optimize this complex system, achieving significant improvements in VR streaming performance and efficiency compared to existing methods.

What Problem Does This Paper Address?

Virtual reality (VR) streaming, especially for high-resolution 360-degree videos, demands extremely high transmission rates and low latency to prevent motion sickness. Existing wireless systems struggle to meet these demands, particularly in multiuser scenarios where multiple users might request overlapping content. Rate-splitting (RS) offers a way to exploit shared content interests by splitting data into common and private messages, but its performance is limited by the user with the weakest signal. Intelligent Reflecting Surfaces (IRS) can enhance signal strength and mitigate interference, but combining RS and IRS for VR streaming and optimizing their joint operation presents a complex, computationally intensive challenge. This paper aims to bridge this gap by developing an efficient and effective system that leverages both RS and IRS for optimal VR streaming performance.

Methods

The core of this work lies in the Deep-GRAIL algorithm, a sophisticated deep reinforcement learning (DRL) framework designed to tackle the complex optimization problem of IRS-aided RS VR streaming. Key methodological contributions include:

  • Rate-Splitting (RS): Used to exploit shared interests among VR users by encoding common content into a shared message and unique content into private messages.
  • Intelligent Reflecting Surfaces (IRS): Deployed to enhance signal propagation and mitigate interference, thereby improving the minimum signal-to-interference-plus-noise ratio (SINR) experienced by users, which is crucial for RS performance.
  • Deep Deterministic Policy Gradient (DDPG) with Imitation Learning: Deep-GRAIL combines the actor-critic DDPG framework with imitation learning. This allows the agent to learn not only from its own exploration but also from the solutions generated by a conventional optimization algorithm (AO), effectively leveraging the hidden convexity of the problem.
  • RavNet: A custom deep neural network architecture designed for the actor network in Deep-GRAIL. It incorporates Differentiable Convex Optimization (DCO) layers to directly handle the convex constraints of the optimization problem during the learning process.
  • q-Step Return: Used in the critic network to improve the accuracy of Q-value estimation and mitigate overestimation issues, leading to more stable learning.
  • Demonstration Replay: A novel replay buffer that stores system transition tuples obtained from an AO algorithm, providing the learning agent with expert demonstrations.

Datasets

The paper utilizes a real-world VR streaming dataset from [13] to model user video tile requests. This dataset captures the head movements of 20 users during multiple VR streaming sessions, providing realistic data for simulating user FoVs and tile requests.

Key Insights and Ideas

  • Synergistic Power of RS and IRS: The combination of RS and IRS is shown to be highly effective for VR streaming, with each technology complementing the other. IRS boosts the minimum SINR, which directly benefits RS by allowing more data to be transmitted via the common message, thus exploiting shared interests more effectively.
  • Imitation Learning for Hidden Convexity: Leveraging imitation learning from an AO algorithm is a powerful strategy to guide DRL agents in complex optimization problems with hidden convexity, leading to more efficient policy learning and better performance.
  • DCO Layers for Constraint Handling: Integrating DCO layers directly into the neural network architecture provides an elegant way to enforce complex constraints during the learning process, a significant advancement in applying DRL to constrained optimization.
  • IRS's Critical Role: The presence of an IRS significantly improves system sum-rate, even with a moderate number of reflecting elements. This highlights the growing importance of IRS technology in future wireless systems.

Key Results

  • The proposed Deep-GRAIL algorithm significantly outperforms several baseline schemes (IRS-aided RS VR with AO, IRS-aided RS VR with SL, IRS-aided RS-NOUM, and IRS-aided MU) in terms of system sum-rate, achievable bitrate, and online execution runtime.
  • For instance, with L=160 reflecting elements, Deep-GRAIL achieves a system sum-rate that is 2.8% to 66.1% higher than the compared methods.
  • The average achievable rate of the common message is shown to increase by 128.3% with the deployment of an IRS (L=100) compared to a system without an IRS.
  • The runtime of Deep-GRAIL is substantially lower than traditional AO-based algorithms, making it more practical for real-time VR streaming applications. For example, it is only 0.36% to 0.55% of the runtime of the AO-based IRS-aided RS VR system.
  • Imitation learning was found to provide the highest performance improvement among the components tested in the ablation study.

Limitations and Future Work

  • Imperfect Channel Estimation: The paper acknowledges that the proposed algorithm assumes perfect CSI. While they investigate the impact of imperfect CSI, further research could explore more robust DRL algorithms that can better handle estimation errors.
  • Subjective Quality of Experience: The current work focuses on objective metrics. Future work could incorporate subjective user experience into the optimization framework.
  • Scalability with Users and IRS Elements: While the learning-based approach shows improved scalability compared to AO, further investigation into the performance and computational complexity with a very large number of users and IRS elements would be beneficial.
  • Extension to Other Applications: The developed framework could potentially be extended to other latency-sensitive and bandwidth-intensive applications beyond VR streaming.

Summary updated

BibTeX
@article{wong2023rate,
  author = {Vincent Wong and Rui Huang and Vincent W. S. Wong and Robert Schober},
  title = {{Rate-Splitting for Intelligent Reflecting Surface-Aided Multiuser VR Streaming}},
  journal = {IEEE JSAC},
  year = {2023},
  doi = {10.1109/jsac.2023.3240710},
}
Figure