Question

Your team is tuning a deep neural network on Vertex AI Vertex AI Hyperparameter Tuning. The search space includes several continuous parameters (learning rate, dropout rate, L2 regularization) plus one discrete parameter (number of hidden layers). Each trial is expensive, taking roughly 6 hours on a GPU, and your total budget allows about 40 trials. You want the tuning service to intelligently use results from completed trials to guide which hyperparameter combinations to try next, minimizing wasted trials. Which search algorithm should you configure?

Accepted Answer

Bayesian optimization (the default Vizier algorithm), which builds a surrogate model from prior trial results to select promising configurations

Answer

Grid search, so every combination in the discrete and continuous ranges is evaluated systematically

Answer

Random search, so trials are sampled independently and can all run in parallel

Answer

Manual search, adjusting hyperparameters yourself between batches of trials based on observed metrics