PRODUCTION SCENARIO
An ML engineer tunes an image classification neural network that publishes an accuracy metric after every epoch. The tuning job has a fixed compute budget, runs many training jobs in parallel, and should stop configurations that are clearly underperforming before they finish.
Which hyperparameter tuning strategy fits these requirements?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Hyperband over the specified ranges
Hyperband uses intermediate and final results to reallocate epochs to configurations performing well and stops those that underperform, which is what a fixed budget and a per-epoch metric allow. Bayesian optimization picks the next job from previous final results, so weak jobs run to completion. Random search ignores previous results, and grid search supports categorical parameters only.