File(s) under permanent embargo
Exploration enhanced expected improvement for Bayesian optimization
conference contribution
posted on 2019-01-01, 00:00 authored by Julian BerkJulian Berk, Tien Vu Nguyen, Sunil GuptaSunil Gupta, Santu RanaSantu Rana, Svetha VenkateshSvetha VenkateshBayesian optimization (BO) is a sample-efficient method for global optimization of expensive, noisy, black-box functions using probabilistic methods. The performance of a BO method depends on its selection strategy through an acquisition function. This must balance improving our understanding of the function in unknown regions (exploration) with locally improving on known promising samples (exploitation). Expected improvement (EI) is one of the most widely used acquisition functions for BO. Unfortunately, it has a tendency to over-exploit, meaning that it can be slow in finding new peaks. We propose a modification to EI that will allow for increased early exploration while providing similar exploitation once the system has been suitably explored. We also prove that our method has a sub-linear convergence rate and test it on a range of functions to compare its performance against the standard EI and other competing methods. Code related to this paper is available at: https://github.com/jmaberk/BO_with_E3I.
History
Event
European Machine Learning and Data Mining. Conference (2018 : Dublin, Ireland)Volume
11052Series
European Machine Learning and Data Mining ConferencePagination
621 - 637Publisher
SpringerLocation
Dublin, IrelandPlace of publication
Cham, SwitzerlandPublisher DOI
Start date
2018-09-10End date
2018-09-14ISSN
0302-9743eISSN
1611-3349ISBN-13
9783030109271Language
engPublication classification
E1 Full written paper - refereedCopyright notice
2019, Springer Nature Switzerland AGEditor/Contributor(s)
M Berlingerio, F Bonchi, T Gärtner, N Hurley, G IfrimTitle of proceedings
ECML-PKDD 2018 : Proceedings of the e European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases 2018Usage metrics
Categories
No categories selectedKeywords
Licence
Exports
RefWorks
BibTeX
Ref. manager
Endnote
DataCite
NLM
DC