Reinforcement learning for ABR
summary
Claims a production ABR replaces heuristics with reinforcement learning trained on real throughput traces and exported from TensorFlow Python into TensorFlow.js for HLS.js integration. Lab tests use Puppeteer and throughput profiles and compare to Pensieve and hybrid baselines. Training focuses on reward shaping; common failure modes are collapsed ladders, high entropy, and dropped frames. Early results match primary QoE metrics in A/B tests, show more quality switches and a noted 50% better performance when running on CPU in lab.
AI-written summary · about
chapters (24)
- 0:00Intro
- 0:30Line up
- 0:50Hotstar
- 1:09Playback QOE
- 1:37Player KPI
- 2:04Adaptive Bitrate Switching
- 2:39ABR Algorithms
- 3:43Data Driven ABR
- 5:17Plan
- 5:33Tasks
- 7:48Pensieve
- 8:27State
- 8:51Action
- 9:49Linear regression
- 11:10Neural Networks
- 11:33Markov Design Process
- 14:03Player integration
- 14:40Tensorflow.js
- 15:21Challenges
- 16:05JS export
- 16:31Dropped Frames
- 17:09Integration issues
- 17:25A/B test results
- 18:51Questions?
related talks
23mMachine Learning for ABR in production
20mReconstructing proprietary video streaming algorithms
03mImproving ABR at Netflix
21mLearning in situ: a randomized experiment in video streaming
13mUltra-High Resolution Playback Visualization and its Applications
11mFree ABR Renditions for User Generated Content Platforms
16mImproving Streaming Experience with Bayesian Optimization, from AB to...
22mProviding better video experiences for the next billion users.