Mowgli: Passively Learned Rate Control for Real-Time Video
summary
Mowgli trains a rate controller entirely offline from production GCC telemetry, using soft actor critic plus conservative Q-learning and distributional critics to avoid dangerous extrapolation, and is implemented on WebRTC. Evaluation on emulated and real cellular traces shows concrete wins: average video bitrates increase 15–39% while video freeze rates drop 60–100%, and performance nearly matches a fully trained online RL policy without any live-user exploration.
AI-written summary · about
chapters (12)
- 0:00Mowgli project intro
- 1:19Rate control in video
- 2:01Reinforcement learning
- 3:27Online RL challenges
- 4:50Offline RL approach
- 6:57Learning from GCC
- 9:45Training framework
- 11:32System workflow
- 12:15Performance on emulated
- 13:39Online RL comparison
- 14:44Real network results
- 15:25Summary and future
related talks
21mLearning in situ: a randomized experiment in video streaming
30mMLDS 2019: Anubhav Shrivastava of Voot at Machine Learning Developers Summit 2019
57mMachine Learning in the Age of IoT
17mContent Aware Encoding for low latency live streaming encoders using deep learning
17mRECL: Responsive Resource-Efficient Continuous Learning for Video Analytics
41mReinforcement Learning and Minecraft
27mUsing Reinforcement Learning to Optimize IAP Offer Recommendations in Mobile Games
32mAI and Minecraft