Scott Clark, Software Engineer, Yelp
summary
MOE is an open source Bayesian global optimizer that models an expensive black‑box objective with a Gaussian Process, fits its kernel hyperparameters analytically, and selects new evaluations by maximizing Expected Improvement to minimize costly A/B or batch runs. It drives traffic allocation with a multi‑arm bandit style loop, tunes hyperparameters for offline jobs, and runs in production at Yelp where it lifts click‑through rates; interfaces include REST, Python, C++, CUDA.
AI-written summary · about
chapters (28)
- 0:00Intro
- 0:25Agenda
- 0:45What is Optimal Learning
- 1:26Multiarm bandits
- 1:55Random variables
- 2:21Why do we care
- 3:18Exploration vs exploitation
- 3:56Naive Policy
- 4:33Example
- 5:04Un unbounded regret
- 5:45Exploration
- 6:07Real world constraints
- 8:12Global optimization
- 8:31Parameters
- 10:02Real World Example
- 11:32Hyperparameter Optimization
- 12:21How does it work
- 13:34Whats a Gaussian Process
- 16:29Optimization
- 16:55Minimize
- 17:32Stop Criteria
- 18:13Cohorts
- 19:42Expensive batch systems
- 20:19Mo
- 20:59Does it work
- 22:15Release date
- 23:06Docker
- 23:45Wrap up
related talks
26mScott Clark, CEO & Co-Founder, SigOpt at MLconf Seattle 2017
26mML Product Experiments at Scale
46mBayesian Optimization Accelerator
24mUsing Bayesian Optimization to Tune Machine Learning Models
21mRise of the Machines: Removing the Human-in-the-Loop
41mOptimizing Machine Learning Models
12mSmartMoE: Efficiently Training Sparsely-Activated Models through Combiningand...
18mAlexandra Johnson, Software Engineer, SigOpt, NYC 2017