Scalable Global Optimization via Local Bayesian Optimization

Authors: David Eriksson, Michael Pearce, Jacob Gardner, Ryan D. Turner, Matthias Poloczek

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental A comprehensive evaluation demonstrates that Tu RBO outperforms stateof-the-art methods from machine learning and operations research on problems spanning reinforcement learning, robotics, and the natural sciences.
Researcher Affiliation Collaboration David Eriksson Uber AI eriksson@uber.com Michael Pearce University of Warwick m.a.l.pearce@warwick.ac.uk Jacob R Gardner Uber AI jake.gardner@uber.com Ryan Turner Uber AI ryan.turner@uber.com Matthias Poloczek Uber AI poloczek@uber.com
Pseudocode No The paper describes the algorithm in prose but does not include a formal pseudocode block or algorithm listing.
Open Source Code Yes An implementation of Tu RBO is available at https://github.com/uber-research/Tu RBO.
Open Datasets Yes We evaluate Tu RBO on a wide range of problems: a 14D robot pushing problem, a 60D rover trajectory planning problem, a 12D cosmological constant estimation problem, a 12D lunar landing reinforcement learning problem, and a 200D synthetic problem. The 14D robot pushing problem is considered in Wang et al. [45]. The rover trajectory planning problem is from [45]. The cosmological constant problem uses a physics simulator1 (https://lambda.gsfc.nasa.gov/toolbox/lrgdr/). The lunar lander is implemented in the Open AI gym2 (https://gym.openai.com/envs/Lunar Lander-v2). The 200-dimensional Ackley function is a standard synthetic problem.
Dataset Splits Yes We choose 20 uniformly distributed hypercubes of (base) side length 0.4, each containing 200 uniformly distributed training points.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models, or memory specifications used for running the experiments.
Software Dependencies No The paper does not list specific software dependencies with version numbers (e.g., Python 3.x, library 1.x).
Experiment Setup Yes We run each method for a total of 10K evaluations and batch size of q = 50. Tu RBO-1 and all other methods are initialized with 100 points except for Tu RBO-20 where we use 50 initial points for each trust region.