Local Latent Space Bayesian Optimization over Structured Inputs

Authors: Natalie Maus, Haydn Jones, Juston Moore, Matt J. Kusner, John Bradshaw, Jacob Gardner

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We apply LOL-BO to six high-dimensional, structured BO tasks over molecules [18, 19] and arithmetic expressions. We additionally investigate the impact of various components of our method in subsection 5.4 and subsection 5.4. and 5 Experiments Figure 1: Optimization results for the log P and arithmetic expressions tasks.
Researcher Affiliation Academia Natalie T. Maus Department of Computer and Information Science University of Pennsylvania Philadelphia, PA nmaus@seas.upenn.edu Haydn T. Jones Los Alamos National Laboratory Los Alamos, NM hjones@lanl.gov Juston S. Moore Los Alamos National Laboratory Los Alamos, NM juston@lanl.gov Matt J. Kusner Centre for Artificial Intelligence, University College London London, UK m.kusner@ucl.ac.uk John Bradshaw Department of Chemical Engineering, Massachusetts Institute of Technology Cambridge, MA jbrad@mit.edu Jacob R. Gardner Department of Computer and Information Science University of Pennsylvania Philadelphia, PA jacobrg@seas.upenn.edu
Pseudocode No The paper does not contain a clearly labeled 'Pseudocode' or 'Algorithm' section or figure.
Open Source Code Yes We implement LOL-BO leveraging Bo Torch [2] and GPy Torch [10]1. Code and model weights are available at https://github.com/nataliemaus/lolbo.
Open Datasets Yes Datasets All methods have access to the same amount of supervised and unsupervised data for each task. To initialize all molecular optimization tasks, we generate labels for a random subset of 10, 000 molecules from the standardized unlabeled dataset of 1.27M molecules from the Guacamol benchmark software [3].
Dataset Splits No The paper states the total size of datasets used (e.g., '10,000 molecules', '40,000 expressions') and that some are labeled, but does not specify the explicit train/validation/test dataset splits (e.g., percentages or absolute counts for each split) used for their experiments.
Hardware Specification Yes All experiments were performed on a single NVIDIA A100 GPU.
Software Dependencies No The paper mentions software frameworks like 'Bo Torch [2]' and 'GPy Torch [10]' but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes Hyperparameters. For the trust region dynamics, all hyperparameters including the initial base and minimum trust region lengths Linit, Lmin, and success and failure thresholds succ, fail are set to the Tu RBO defaults as used in Eriksson et al. [8]. Our method introduces only one new hyperparameter, retrain, which we set to 10 in all experiments. and Because of the small evaluation budget typically considered for these benchmarks, we use a batch size of 1.