High-Dimensional Discrete Bayesian Optimization with Intrinsic Dimension

  • Shu Jun Li
  • , Mingjia Li
  • , Hong Qian*
  • *Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Bayesian optimization (BO) has achieved remarkable success in optimizing low-dimensional continuous problems. Recently, BO in high-dimensional discrete solution space is in demand. However, satisfying BO algorithms tailored to this issue still lack. Fortunately, it is observed that high-dimensional discrete optimization problems may exist low-dimensional intrinsic subspace. Inspired by this observation, this paper proposes a Locality Sensitive Hashing based Bayesian Optimization (LSH-BO) method for high-dimensional discrete functions with intrinsic dimension. Via randomly embedding solutions from intrinsic subspace to original space and discretization, LSH-BO turns high-dimensional discrete optimization problems into low-dimensional continuous ones. Theoretically we prove that, with probability 1, there exists a corresponding optimal solution in the intrinsic subspace. The empirically results on both synthetic functions and binary quadratic programming task verify that LSH-BO surpasses the compared methods and possesses the versatility across low-dimensional and high-dimensional kernels.

Original languageEnglish
Title of host publicationPRICAI 2022
Subtitle of host publicationTrends in Artificial Intelligence - 19th Pacific Rim International Conference on Artificial Intelligence, PRICAI 2022, Proceedings
EditorsSankalp Khanna, Jian Cao, Quan Bai, Guandong Xu
PublisherSpringer Science and Business Media Deutschland GmbH
Pages534-547
Number of pages14
ISBN (Print)9783031208614
DOIs
StatePublished - 2022
Event19th Pacific Rim International Conference on Artificial Intelligence, PRICAI 2022 - Shangai, China
Duration: 10 Nov 202213 Nov 2022

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume13629 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference19th Pacific Rim International Conference on Artificial Intelligence, PRICAI 2022
Country/TerritoryChina
CityShangai
Period10/11/2213/11/22

Keywords

  • Black-box optimization
  • Intrinsic subspace
  • Locality sensitive hashing

Fingerprint

Dive into the research topics of 'High-Dimensional Discrete Bayesian Optimization with Intrinsic Dimension'. Together they form a unique fingerprint.

Cite this