Question

A data scientist has already run a SageMaker automatic model tuning (AMT) job for an XGBoost classifier and found a strong region of the hyperparameter space. The team now wants to expand the search by adding two new hyperparameters and widening some existing ranges, but they want the new tuning job to leverage the knowledge gained from the completed job so it converges faster and uses fewer training jobs. Which approach best meets this requirement?

Accepted Answer

Create a new tuning job configured as a warm start job of type TRANSFER_LEARNING that references the previous tuning job as a parent.

Answer

Create a new tuning job configured as a warm start job of type IDENTICAL_DATA_AND_ALGORITHM that references the previous tuning job as a parent.

Answer

Re-run the original tuning job with a higher MaxNumberOfTrainingJobs value and the Bayesian strategy so it reuses prior results automatically.

Answer

Export the best hyperparameters from the previous job and hardcode them as static values, then launch a Random search over the two new hyperparameters only.