Bayesian optimization uses a surrogate model to guide hyperparameter search. How does it differ from random search?