Bayesian Optimization with High-Dimensional Outputs

Authors: Wesley J. Maddox, Maximilian Balandat, Andrew G. Wilson, Eytan Bakshy

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate empirically how large-scale sampling from MTGPs can aid in challenging multi-objective, constrained, and contextual Bayesian Optimization problems (Section 4).
Researcher Affiliation Collaboration Wesley J. Maddox New York University wjm363@nyu.edu Maximilian Balandat Facebook balandat@fb.com Andrew Gordon Wilson New York University andrewgw@cims.nyu.edu Eytan Bakshy Facebook eytan@fb.com
Pseudocode No The paper describes methods and procedures in narrative text and mathematical equations, but does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes Our code is fully integrated into Bo Torch, see https://botorch.org/tutorials/composite_ bo_with_hogp and https://botorch.org/tutorials/composite_mtbo for tutorials.
Open Datasets Yes We consider a multi-task version of the Hartmann-6 function... We compare Matheron sampled MTGPs to batch independent MTGPs on the C2DTLZ2 [19]... and OSY [43]... Lunar Lander... from the Open AI Gym [8]... MOPTA08 benchmark problem [35]... Chemical Pollutants... originally defined in Bliznyuk et al. [5]... Optimizing PDEs... solved in py-pde [65]... Cell-Tower Coverage: Following Dreifuerst et al. [21]... Optical Interferometer... as in Sorokin et al. [53].
Dataset Splits No The paper describes experiments on black-box optimization functions and simulated environments where data points are queried sequentially, rather than using pre-existing datasets with explicit train/validation/test splits.
Hardware Specification Yes on a single Tesla V100 GPU (a,b) and on a single CPU (c,d).
Software Dependencies No The paper mentions software like 'Bo Torch' and 'py-pde' but does not provide specific version numbers for these or other ancillary software components.
Experiment Setup Yes Following Daulton et al. [17] we use both q Par EGO and q EHVI with q = 2, for C2DTLZ2 and optimize for 200 iterations... using a batch size of q = 10, optimizing for 30 iterations... initialize with 150 data points and use Tu RBO with Thompson sampling with batches of q = 20 for a total of 1000 function evaluations and repeat over 30 trials.