Learning search spaces for Bayesian optimization: Another view of hyperparameter transfer learning