Improving Streaming Experience with Bayesian Optimization, from AB to...
summary
Netflix claims that Bayesian optimization plugged into its A/B experimentation framework finds optimal ABR settings across play delay quality and rebuffer trade-offs by building surrogate models and iterating a few experiments. The pipeline turns a combinatorial space (ten values for three knobs is 1,000 combinations; 100 values is a million) into focused exploration near the Pareto frontier, needs only a couple of rounds of A/B tests in practice, and materially reduces search time.
AI-written summary · about
chapters (12)
- 0:00Intro
- 1:18What would be the perfect startup QoE?
- 2:18ABR to the rescue!
- 2:48Challenge #1: Simultaneously optimize all QoE metrics under constrained resources
- 3:49Our job is to find the optimal operating points (pareto frontier)
- 5:24How to address these 2 challenges?
- 7:18How to find the next candidates that further minimize rebuffers?
- 8:08Ask a first grader!
- 9:01Find the range that minimizes the probability of rebuffer
- 10:30Bayesian Optimization scales well
- 11:44Iteratively increase the precision of pareto frontier
- 14:43Summary Bayesian Optimization is an efficient exploration methodology
related talks
03mImproving ABR at Netflix
37mRealtime Bayesian A-B testing with Spark Streaming
21mLearning in situ: a randomized experiment in video streaming
25mScaling Bayesian Experimentation
22mProviding better video experiences for the next billion users.
28mBayesian Optimization- Fundamentals, Implementation, and Practice
46mBayesian Optimization Accelerator
38mUsable A/B testing – A Bayesian approach