Learning in situ: a randomized experiment in video streaming
summary
A randomized, blinded deployment streams 38.6 years of video to 63,508 users and shows huge statistical uncertainty: single-day experiments produce overlapping 95% CIs and roughly two years per scheme are needed to detect ~20% differences. A simple buffer-based ABR (BBA) often matches or beats complex and learned policies in real Internet conditions. A new system, Fugu, uses model-predictive control plus a learned transmission-time predictor that outputs a probability distribution and consumes TCP kernel stats; trained in situ, it robustly outperforms BBA.
AI-written summary · about
chapters (5)
- 0:00Introduction and objectives
- 1:45Uncertainty in ABR metrics
- 8:32Simple buffer-based control
- 11:43Learning in situ with Fugu
- 16:05Conclusion and Q&A
related talks
20mReconstructing proprietary video streaming algorithms
23mMachine Learning for ABR in production
16mImproving Streaming Experience with Bayesian Optimization, from AB to...
30mMLDS 2019: Anubhav Shrivastava of Voot at Machine Learning Developers Summit 2019
13mSENSEI: Aligning Video Streaming Quality with Dynamic User Sensitivity
19mReinforcement learning for ABR
17mYour Player is Smarter Than You Think: Self-Learning in Adaptive Streaming
26mMagic Mirror: Understanding the Diversified Network Conditions of Twitch Community