FairEdge360: Distributed Multi-Agent Reinforcement Learning for QoE-Fair 360° Video Streaming with Uncertainty-Aware...

Area of Science:

• Computer Science
• Networking
• Artificial Intelligence

Background:

• Shared immersive environments require simultaneous 360° video streaming from multiple users over edge networks.
• Existing adaptive bitrate systems optimize viewers individually, leading to bandwidth competition and poor fairness (Jain's Index < 0.85).

Purpose of the Study:

• To develop a hierarchical multi-agent reinforcement learning framework, FairEdge360, for equitable and high-quality multi-user 360° video streaming.
• To demonstrate that fairness and quality are complementary objectives in this context.

Main Methods:

• Reformulated multi-user streaming as a Decentralised Partially Observable Markov Decision Process (Dec-POMDP).
• Introduced a Lightweight Uncertainty Estimator (LUE) for efficient per-device viewport prediction confidence evaluation.
• Utilized a variational Graph Neural Network for state compression and a dynamic neighborhood graph for communication.
• Implemented an edge coordinator maximizing Nash social welfare (NSW) and employed counterfactual advantage estimation for credit assignment.

Main Results:

• FairEdge360 improved Jain's Fairness Index from 0.934 to 0.976 (+4.5%).
• Worst-case user quality-of-experience (MOS) increased from 2.54 to 3.21 (+26.4%).
• Rebuffering rate was halved from 2.1% to 1.1%, with reduced device energy consumption (38.9%) and communication overhead (75%).

Conclusions:

• FairEdge360 effectively addresses the fairness-quality trade-off in multi-user 360° streaming.
• The framework significantly enhances user experience and network efficiency in shared immersive environments.
• The proposed methods, including LUE and NSW maximization, are crucial for achieving these improvements.