EDQD: An Edge-Driven Multi-Agent DRL Solution for Improving Joint QoE in DASH-based Rich Media Content Delivery

In the realm of adaptive video streaming, ensuring high Quality of Experience (QoE) across multiple clients in heterogeneous network environments stands as a significant challenge. This paper delves into the augmentation of QoE fairness and adaptive bitrate aggregation in Dynamic Adaptive Streaming over HTTP (DASH), particularly tailored for multiple rich media clients. This study proposes an edge-driven MultiAgent Deep Reinforcement Learning (MADRL) solution for improving joint QoE in DASH-based rich media content delivery (EDQD) to orchestrate QoE-centric video streaming framework. In order to address the complexities of network conditions at the edge for optimal media streaming bitrate allocation the cooperative strategies among agents are of great importance. Therefore, the EDQD objective is to training agents in such a way to learn policies that optimize not only individual client QoE but also ensure fairness among clients to enhance the overall rich streaming experience. Here, in comparison with previously utilized Asynchronous Advantage Actor-Critic (A3C) algorithm, the Counterfactual Multi-Agent Policy Gradients (COMA) experimentation and training based on 5G network traces has the potential to demonstrate substantial improvements in joint $\mathbf{Q o E}$ by efficiently allocating bitrate across multiple client in the edgedriven scenario. This research provides a major contribution to the emerging field of QoE-driven rich video streaming by highlighting the effectiveness of novel MADRL-based approach in edge-driven architectures that is designed to provide immersive and high-fidelity rich media experiences. In terms of average QoE, the experimental findings reported in this work outperform other state-of-the-art adaptive edge driven bitrate algorithms by $6.9 \%$ and $16.8 \%$.