A team can afford only twelve trials in a large continuous hyperparameter space. Which approach avoids enumerating every point of a fixed grid?

1. Scoring every trial only on the training examples
2. Selecting the parameter name with the fewest characters
3. A full exhaustive search over every possible real number
4. Random search over specified distributions