GCP Professional ML Engineer Practice Question 5: Vizier Black Box Optimization
Google Cloud Professional Machine Learning Engineer · Test 1 · question 5 of 10
An ML team trains a large model on an on-premises GPU cluster that, for data residency reasons, cannot be moved to Google Cloud. Each training run takes about nine hours, so exhaustive search is impractical, and the team wants Google's managed optimization service to propose successive hyperparameter values based on the results already observed.
Which service should the team use?
Show answer and explanation
Answer:
Agent Platform Vizier (formerly Vertex AI Vizier) is a black-box optimization service: it optimizes any system with configurable parameters where each evaluation is expensive, and it explicitly works with any system you can evaluate; the training code does not have to run on Google Cloud. The interaction model is a study whose trials your own trainer requests and whose measurements your trainer reports back, which is precisely what a hybrid setup like this needs. Nine-hour evaluations are the classic case for sequential, model-based suggestion rather than brute force.
Option A confuses the two tuning surfaces. A hyperparameter tuning job is a custom training job with a tuning configuration wrapped around it; the trials run on Agent Platform's own compute. There is no mechanism to point a tuning job's worker pool at an on-premises cluster.
Option C is not a transfer that works. AutoML's internal architecture and hyperparameters are tied to the models it builds; there is nothing meaningful to lift out and paste into a different model trained on different infrastructure.
Option D delivers grid search, which is the strategy the stem rules out. A grid over a nine-hour objective burns the budget uniformly instead of concentrating trials where results are promising, and Experiments only records runs; it does not suggest the next set of values.
Reference: Agent Platform Vizier overview