Advanced
Open
Pro
What Must the Off-Policy Gate Show Before Launch?
Part of the ML System Design Interview path →
Part of the Reinforcement Learning & Long-term Optimization path →
Part of the Advanced Data Science & Experimentation path →
You have a candidate surge policy trained on two weeks of logs from the incumbent rule with 20 percent uniform exploration. Offline, its estimated value is higher than the rule's. A colleague says "ship it to the switchback".
- List what the logs must contain for off-policy evaluation to be possible at all.
- List what the gate must show, beyond a higher point estimate.
- Explain what the gate cannot tell you, and what you do about that.
Share this question