PMLE Scaling Prototypes into ML Models Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✓

  Bayesian optimization using Vertex AI Vizier.

  Why this is correct

  Bayesian optimisation builds a probabilistic surrogate model of validation accuracy and selects each configuration to maximise expected improvement, converging in far fewer trials than grid or random search. This suits the 50-job budget while targeting the stated objective metric.

• ✗

  Random search with 50 random configurations.

  Why it's wrong here

  Random search samples configurations independently, ignoring results from completed trials, so it cannot concentrate the 50-job budget around promising regions of the search space. It suits cheap, high-dimensional tuning where exhaustive coverage matters. Vertex AI Vizier's Bayesian optimisation uses prior trial outcomes to guide sampling, which is what efficient tuning under a fixed budget requires.

• ✗

  Use a custom algorithm implemented in the training code.

  Why it's wrong here

  Vizier runs its own Bayesian optimisation loop; a custom algorithm inside the training code cannot feed trial results back to the Vizier service, so the 50-job budget is spent without guided search. Custom logic suits bespoke pipelines where no managed tuning service exists.

• ✗

  Grid search with 50 evenly spaced points.

  Why it's wrong here

  Grid search evaluates a fixed lattice, so with 50 jobs across several hyperparameters most dimensions get only a few values and promising regions between grid points are missed. It suits very low-dimensional spaces where exhaustive coverage is affordable.

Go deeper

Related to this question