Reconstructing proprietary video streaming algorithms
summary
Reconstructs proprietary ABR logic from player and network logs across roughly 9 hours of traces, learning interpretable concise decision trees over about 1,300 primitives. Produces high‑accuracy reconstructions for 7 of 10 platforms with trees of ≤20 rules, compares against imitation‑learning and rate baselines, and measures agreement plus perceptual similarity; nine reconstructions match user‑experience metrics within 6% on average. Fails where decision space differs (YouTube segmenting, Pornhub progressive download).
AI-written summary · about
chapters (20)
- 0:00Intro
- 0:17Roadmap
- 0:43Streaming in Industry
- 1:09Streaming in academia
- 1:26Applications of reconstruction
- 3:00Introduction into ABRs
- 3:38Simple ABR in code
- 4:02Simple ABR as a tree
- 4:39Why do we need primitives
- 6:20Number of primitives
- 6:41Services of interest
- 7:11Inference features
- 7:52Why is agreement not sufficient
- 11:15General metric analysis
- 13:29Perceptual metric analysis
- 15:22Comparing industry with academia
- 16:22Analysing the simplified SRF tree
- 17:25What usage has interpretability
- 17:57Limitations & Future Work
- 19:15Takeaways
related talks
21mLearning in situ: a randomized experiment in video streaming
19mReinforcement learning for ABR
23mMachine Learning for ABR in production
26mMagic Mirror: Understanding the Diversified Network Conditions of Twitch Community
30mMLDS 2019: Anubhav Shrivastava of Voot at Machine Learning Developers Summit 2019
21mWhat to do after per-title encoding
17mYour Player is Smarter Than You Think: Self-Learning in Adaptive Streaming
16mImproving Streaming Experience with Bayesian Optimization, from AB to...