Efficient Hyperparameter Optimization and Infinitely Many Armed Bandits

Performance of machine learning algorithms depends critically on identifying a good set of hyperparameters. While current methods offer efficiencies by adaptively choosing new configurations to train, an alternative strategy is to adaptively allocate resources across the selected configurations. We formulate hyperparameter optimization as a pure-exploration non-stochastic infinitely many armed bandit problem where allocation of additional resources to an arm corresponds to training a configuration on larger subsets of the data. We introduce Hyperband for this framework and analyze its theoretical properties, providing several desirable guarantees. We compare Hyperband with state-of-the-art Bayesian optimization methods and a random search baseline on a comprehensive benchmark including 117 datasets. Our results on this benchmark demonstrate that while Bayesian optimization methods do not outperform random search trained for twice as long, Hyperband in favorable settings offers valuable speedups.