The Players and Playback Working Group meeting featured a presentation by Abdelhak on CADENCE, a multi-agent deep reinforcement learning content steering system for CDN selection. Abdelhak explained how CADENCE improves video streaming quality of experience while reducing CDN costs through decentralized CDN switching that can adapt to dynamic network conditions. The system uses a centralized training approach with decentralized inference, collecting telemetry data from multiple CDNs including those not currently being used, and employs active probing to maintain fresh performance metrics. The presentation included results from testing CADENCE in both video-on-demand and low-latency streaming modes, demonstrating superior performance compared to existing static and dynamic CDN switching methods while scaling to hundreds of clients in controlled environments. Yuriy suggested exploring more deterministic optimization approaches and formal verification methods, which Abdelhak confirmed they are working on through a Mixed Integer Linear Programming model. The group discussed potential collaboration opportunities, including adding references to Abdelhak’s work in an upcoming report on multi-CDN switching and creating a centralized benchmarking platform for CDN selection algorithms.
1. Abdelhak – Send the presentation slides to Daniel for distribution to the group.
2. Abdelhak – Review the upcoming SVTA report on multi-CDN switching (as invited by Yuriy) and consider providing feedback and references.
3. Daniel – Distribute the recording and presentation slides to the group (after receiving them from Abdelhak).
4. Daniel – Check with Lindsey (or relevant person) about the possibility for working group chairs to send emails via the SVTA email reflector or alternative distribution methods.
5. Yuriy – Update the group once the content steering framework website is rebuilt and available.
- Multi-CDN Selection and Implementation
- The Players and Playback Working Group meeting featured a presentation by Abdelhak on a recent paper accepted in MMCS titled “Cadence,” which focuses on improving quality of experience and reducing CDN costs through multi-CDN selection.
- Abdelhak explained the benefits of multi-CDN approaches compared to single-CDN systems, citing real-world examples of outages at Cloudflare and other CDNs.
- The presentation covered end-to-end video streaming pipelines and addressed common challenges with multi-CDN implementations.
CDN Switching Learning Approach
- Abdelhak discussed the challenges of switching between multiple CDN providers and presented their work on developing a learning-based approach to efficiently switch between CDNs while maintaining quality of experience and reducing costs.
- He explained the evolution from single-agent to multi-agent reinforcement learning to address scalability issues.
- Abdelhak also reviewed different CDN switching implementation methods, including DNS-based, manifest-based, server-side, and client-side approaches, before introducing content steering as a centralized solution.
- He concluded by highlighting the limitations of static CDN switching methods, such as round-robin and weighted round-robin, in handling CDN degradation or outages.
CADENCE CDN Switching Solution
- Abdelhak presented a new CDN switching solution called CADENCE, which uses multi-agent deep reinforcement learning to address limitations in existing static and dynamic CDN switching approaches.
- He explained that CADENCE aims to maximize quality of experience while reducing costs and preventing overload, using a dual-objective approach with centralized training but decentralized inference.
- When Yasser asked about the softmax module, Abdelhak described it as the final layer that produces a vector of probabilities to determine the best CDN selection based on quality experience and cost considerations.
CDN Selection System Architecture Discussion
- Abdelhak explained how the softmax function in their system selects the best CDN based on probabilities calculated from actor network vectors, with the probabilities summing to 1.
- When asked about training the neural networks, Abdelhak described using an emulator tool that simulates 4 CDNs and 1,000 clients across different network types (5G, 4G, satellite) to collect bandwidth traces over a month or two.
- Abdelhak also mentioned they are working on concept drift solutions and exploring hybrid approaches combining rule-based and learning-based methods.
CDN Verification and Telemetry Models
- Abdelhak discussed his team’s work on deterministic models from SMT solvers to verify non-deterministic model outputs, which received positive feedback from industry collaborators regarding verification needs.
- He explained their current telemetry system that collects throughput, rebuffering, and delays per CDN, along with an active probing system to maintain data freshness when clients don’t connect to all CDNs.
- The reward function combines these metrics with costs and penalizes overutilization of the same CDN to determine the best CDN for each client.
- The discussion was interrupted by Daniel’s question about how audio and video segments from the same CDN are handled in the 10-sample collection process.
CDN Selection System with Reinforcement Learning
- Abdelhak presented CADENS, a content delivery network (CDN) selection system that uses reinforcement learning to optimize video streaming performance while considering cost.
- The system was tested in both video-on-demand and low-latency streaming modes, demonstrating better VMAF quality and cost efficiency compared to existing approaches.
- The team is working on a Mixed Integer Linear Programming (MILP) model as a more scalable alternative to the current learning-based approach, and they are developing a benchmarking platform for multi-CDN solutions.
- Yuriy mentioned an upcoming report on multi-CDN switching that Abdelhak’s work could help support with additional references.
Next Meeting: Monday, May 25, 2026 8am PDT
Chat Record: