Question 13
Domain 2: Explore data and run experimentsYou tune hyperparameters on your HyperDrive experiment using Random sampling. You want to terminate 30 percent of the lowest performing runs at each evaluation interval, based on the primary metric. Which early termination policy should you use?
Correct answer: D
Explanation
The Truncation Selection policy fits because it is designed to stop a fixed fraction of the worst runs at each evaluation point, matching “terminate 30 percent of the lowest performing runs.” It uses the primary metric and supports a `truncation_percentage` setting, unlike Bandit or Median Stopping policies.
Why each option is right or wrong
A. Median stopping policy
B. Bandit policy
C. No termination policy
D. Truncation selection policy **Correct answer:** D. Truncation selection policy **Explanation:** The Truncation Selection policy is designed to drop a fixed percentage of the worst‑performing trials at each evaluation checkpoint, exactly matching the requirement to terminate 30 % of runs. It supports parameters such as `truncation_percentage`, `evaluation_interval`, and `delay_evaluation`. In contrast, the Bandit policy uses a slack factor rather than a fixed percentage, the Median Stopping policy cancels runs below the median without a configurable percentage, and choosing no termination policy would allow all runs to complete regardless of performance. **Reference:** - [HyperDrive early termination policies](https://learn.microsoft.com/en-us/azure/machine-learning/how-to-tune-hyperparameters#early-termination-policies)
Azure Machine Learning’s HyperDrive truncation selection policy is the only early-termination policy that lets you specify a fixed `truncation_percentage`, so setting it to 30% will stop the bottom 30% of runs at each `evaluation_interval` using the primary metric. The policy is documented in the HyperDrive early termination policies guidance, which distinguishes it from Bandit (slack-based) and Median Stopping (median-based, not percentage-based).