MLA-C01 Practice Question 45: Bayesian Tuning Parallelism
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 5 · question 5 of 10
An engineer runs a SageMaker AI automatic model tuning job for an XGBoost model. Each training job takes about 10 minutes. The tuning job uses the Bayesian strategy over four continuous hyperparameters, with MaxNumberOfTrainingJobs set to 60 and MaxParallelTrainingJobs also set to 60 so that it finishes quickly. The best objective value it found is no better than a random search of the same size produced last month. The budget stays at 60 training jobs, but the tuning job may now take several hours.
Which change is most likely to find better hyperparameters within the same budget?
Show answer and explanation
Answer:
Bayesian optimization uses what earlier training jobs found to choose better hyperparameters for later ones, so it is sequential by nature and cannot scale massively in parallel. With 60 parallel jobs out of 60 total, every configuration is chosen at the start, before a single result exists. The search collapses into random sampling, which is exactly why it matched last month's random search. Lowering MaxParallelTrainingJobs lets the strategy use prior results, trading the wall-clock time the team can now afford for better choices from the same 60 jobs (option C).
- A is wrong because Hyperband targets large, long-running jobs, where stopping weak configurations early saves meaningful compute. AWS recommends random search or Bayesian optimization for short jobs like these 10-minute runs, and the real flaw, that no job learns from another, remains. - B is wrong because grid search supports only categorical hyperparameters, and these four are continuous. Even with discretized ranges, an exhaustive grid is a poor use of a fixed budget of 60 jobs. - D is wrong because faster individual jobs reduce wall-clock time, which is no longer the constraint. With parallelism still at 60, every configuration is still picked blind.
Reference: Best Practices for Hyperparameter Tuning