GCP Professional ML Engineer Practice Question 13: Experiment Tracking Run Comparison
Google Cloud Professional Machine Learning Engineer · Test 2 · question 3 of 10
Four data scientists are each tuning the same demand model, recording learning rate, tree depth, feature set and validation RMSE in personal spreadsheets. Nobody can reconstruct which combination produced last month's best score, and two people have already re-run the same configuration. You need one managed place where every run's parameters, metrics and produced artifacts are recorded automatically from the training code and can be compared side by side.
Which Google Cloud capability should you adopt?
Show answer and explanation
Answer:
Experiments on Gemini Enterprise Agent Platform is the managed experiment tracker for exactly this problem. It tracks the steps of an experiment run, its inputs (algorithm, parameters, datasets) and its outputs (models, checkpoints, metrics), records the lineage through ML Metadata, and lets you group and compare multiple runs to pick the best model. The SDK logs from inside the training code, so nothing depends on someone remembering to update a sheet, and experiment runs themselves incur no extra charge (option D).
- A is wrong because it is a hand-built tracker: someone must write and maintain the schema, the insert calls and the dashboard, and it still captures no artifact or lineage links back to the model produced by a run. - B is wrong because Cloud Monitoring is built for time series of an operational system's health, not for comparing discrete training runs by their hyperparameters. It has no notion of a run, a parameter set or a model artifact. - C is wrong because a Cloud Storage prefix is just file storage. Comparing runs means opening and diffing JSON by hand, and nothing enforces that two people name or structure their files the same way.
Reference: Introduction to Gemini Enterprise Agent Platform Experiments