Abstract

Multi-Content Delivery Network (Multi-CDN) strategies are vital to enhancing Quality of Experience (QoE) in adaptive video streaming. The recent content steering standard (ETSI TS 103
998) enables real-time CDN selection by collecting performance statistics from players and CDNs. However, existing rule-based approaches such as round-robin, least connections remain static and often fail to adapt to network dynamics.
To address this, we developed StreamWise, a single-agent learning-based framework for CDN selection. Although StreamWise improved QoE, its coarse-grained design overlooked global oper�ational costs, limiting its effectiveness. Building on this, we propose Cadence, a multi-agent deep reinforcement learning framework that jointly optimizes user QoE and multi-CDN costs. Cadence adopts a Centralized Training with Decentralized Execution (CTDE) paradigm, where per-client agents make fine-grained CDN selections, while a centralized critic coordinates training. Both frameworks are trained on experience trajectories from Pensieve ABR, ensuring realistic adaptation
to network and content dynamics.
Through extensive trace-driven emulation experiments, we show that StreamWise improves the average VMAF by 8.5% and achieves a 1.5× higher bitrate, and delivers a 48% QoE improvement over heuristic baselines. Cadence further improves performance, improving VMAF by up to 21%,
achieving 1.2× higher bitrate for live streaming, reducing rebuffering events by up to 10×, and lowering multi-CDN operational costs by 35%.
This thesis demonstrates that reinforcement learning-based Multi-CDN frameworks—first with StreamWise and more effectively with Cadence—deliver high-quality, cost-efficient adaptive video
streaming at scale, significantly outperforming heuristic and coarse-grained approaches.