Abstract
Virtual reality (VR) is promising to fundamentally transform a broad spectrum of industry sectors and the way humans interact with virtual content. However, despite unprecedented progress, current networking and computing infrastructures are incompetent to unlock VR's full potential. In this paper, we consider delivering the wireless multi-tile VR video service over a mobile edge computing (MEC) network. The primary goal is to minimize the system latency/energy consumption and to arrive at a tradeoff thereof. To this end, we first cast the time-varying view popularity as a model-free Markov chain to effectively capture its dynamic characteristics. After jointly assessing the caching and computing capacities on both the MEC server and the VR playback device, a hybrid policy is then implemented to coordinate the dynamic caching replacement and the deterministic offloading, so as to fully utilize the system resources. The underlying multi-objective problem is reformulated as a partially observable Markov decision process, and a deep deterministic policy gradient algorithm is proposed to iteratively learn its solution, where a long short-term memory neural network is embedded to continuously predict the dynamics of the unobservable popularity. Simulation results demonstrate the superiority of the proposed scheme in achieving a trade-off between the energy efficiency and the latency reduction over the baseline methods.
Introduction
Virtual Reality (VR) is a resource-intensive technology that demands ultra-high-resolution video delivery with minimal "motion-to-photon" latency to prevent motion sickness. Mobile VR headsets, while portable, are constrained by limited battery life and processing power. Mobile Edge Computing (MEC) offers a solution by offloading heavy computational tasks, such as 2D-to-3D projection, to servers located near the user.
However, two major hurdles remain: the dynamic and often unpredictable nature of user viewing patterns (Field of View or FoV), and the complex interaction between content caching and task offloading. Most existing models assume static viewing popularity, which fails in real-world scenarios. Furthermore, caching and offloading are typically treated as separate problems, ignoring their deep interdependencies.
System Architecture and Methodology
The authors propose a multi-tile deterministic offloading and dynamic caching scheme. By slicing VR frames into tiles, the system can manage resources at a granular level, processing some tiles at the MEC server while others are handled locally by the headset.
Hybrid Learning Framework
To address the challenges of dynamic popularity and resource interaction, the study introduces a hybrid policy learning framework:
- Model-Free Markov Modeling: The time-varying popularity of video tiles is modeled as a Markov chain following a Zipf distribution, assuming no prior knowledge of transition probabilities.
- LSTM-DDPG Architecture: A Long Short-Term Memory (LSTM) network tracks historical user requests to predict future viewpoint popularity, addressing the partially observable nature of the environment. This prediction is fed into a Deep Deterministic Policy Gradient (DDPG) algorithm to optimize the continuous action space of caching and offloading.
Key Insights
A central insight of this work is that caching and offloading form a feedback loop: current caching status dictates offloading efficiency, while offloading decisions determine future cache contents. Solving these jointly is significantly more efficient than isolated optimization. Additionally, the integration of an LSTM allows the system to pre-cache tiles proactively based on learned behavioral trends. The move to tile-level granularity further maximizes the utilization of idle computing power across the network.
Performance Evaluation
The proposed LSTM-DDPG scheme was evaluated using simulated user traces and standard wireless parameters (802.11ad mmWave). Key results include:
- Latency Reduction: The system successfully converged to an average service latency of approximately 9ms, well within high-quality VR requirements.
- Superior Tradeoff: The hybrid approach consistently outperformed random allocation, traditional DDPG (without LSTM), and Normalized Advantage Functions (NAF) in terms of energy efficiency and latency reduction.
- Flexibility: By adjusting a weight parameter ($\omega$), the system can be tuned to prioritize either extreme energy saving or ultra-low latency.
Conclusion
This research presents a robust framework for optimizing MEC-assisted VR services. By combining predictive LSTM networks with DDPG, the system effectively manages the energy-latency tradeoff in dynamic environments. Future work will focus on multi-user scenarios, complex interference management, and more sophisticated power modeling.
Summary updated