arXiv (Cornell University) · 2019 · 144 citations · 0 references
Artificial IntelligenceLocal Bayesian OptimizationModel OptimizationLarge-scale Global OptimizationEngineeringMachine LearningData ScienceBayesian OptimizationStochastic OptimizationContinuous OptimizationExploration V ExploitationLarge Scale OptimizationBayesian MethodsComputer ScienceStatistical InferenceGlobal Acquisition
Bayesian optimization is sample‑efficient for expensive black‑box functions, yet it struggles with high‑dimensional problems involving thousands of observations and is often outperformed by other paradigms. The authors attribute this limitation to the implicit homogeneity of global probabilistic models and the excessive exploration induced by global acquisition. They introduce TuRBO, a local probabilistic method that fits multiple local models and allocates samples globally through an implicit bandit strategy. Extensive experiments show TuRBO surpasses state‑of‑the‑art machine‑learning and operations‑research methods across reinforcement‑learning, robotics, and natural‑science benchmarks.
Bayesian optimization has recently emerged as a popular method for the sample-efficient optimization of expensive black-box functions. However, the application to high-dimensional problems with several thousand observations remains challenging, and on difficult problems Bayesian optimization is often not competitive with other paradigms. In this paper we take the view that this is due to the implicit homogeneity of the global probabilistic models and an overemphasized exploration that results from global acquisition. This motivates the design of a local probabilistic approach for global optimization of large-scale high-dimensional problems. We propose the $\texttt{TuRBO}$ algorithm that fits a collection of local models and performs a principled global allocation of samples across these models via an implicit bandit approach. A comprehensive evaluation demonstrates that $\texttt{TuRBO}$ outperforms state-of-the-art methods from machine learning and operations research on problems spanning reinforcement learning, robotics, and the natural sciences.