Beyond Local Nash Equilibria for Adversarial Networks

Frans A. Oliehoek; Rahul Savani; Jose Gallego; Elise van der Pol; Roderich Groß

doi:10.1007/978-3-030-31978-6_7

Beyond Local Nash Equilibria for Adversarial Networks

Frans A. Oliehoek^*, Rahul Savani, Jose Gallego, Elise van der Pol, Roderich Groß

^*Corresponding author for this work

Interactive Intelligence

Research output: Chapter in Book/Conference proceedings/Edited volume › Conference contribution › Scientific › peer-review

7 Citations (Scopus)

119 Downloads (Pure)

Abstract

Save for some special cases, current training methods for Generative Adversarial Networks (GANs) are at best guaranteed to converge to a ‘local Nash equilibrium’ (LNE). Such LNEs, however, can be arbitrarily far from an actual Nash equilibrium (NE), which implies that there are no guarantees on the quality of the found generator or classifier. This paper proposes to model GANs explicitly as finite games in mixed strategies, thereby ensuring that every LNE is an NE. We use the Parallel Nash Memory as a solution method, which is proven to monotonically converge to a resource-bounded Nash equilibrium. We empirically demonstrate that our method is less prone to typical GAN problems such as mode collapse and produces solutions that are less exploitable than those produced by GANs and MGANs.

Original language	English
Title of host publication	Artificial Intelligence
Subtitle of host publication	30th Benelux Conference, BNAIC 2018, Revised Selected Papers
Editors	Martin Atzmueller, Wouter Duivesteijn
Place of Publication	Cham
Publisher	Springer
Pages	73-89
Number of pages	17
ISBN (Electronic)	978-3-030-31978-6
ISBN (Print)	978-3-030-31977-9
DOIs	https://doi.org/10.1007/978-3-030-31978-6_7
Publication status	Published - 2019
Event	30th Benelux Conference on Artificial Intelligence, BNAIC 2018 - ‘s-Hertogenbosch, Netherlands Duration: 8 Nov 2018 → 9 Nov 2018

Publication series

Name	Communications in Computer and Information Science
Volume	1021
ISSN (Print)	1865-0929
ISSN (Electronic)	1865-0937

Conference

Conference	30th Benelux Conference on Artificial Intelligence, BNAIC 2018
Country/Territory	Netherlands
City	‘s-Hertogenbosch
Period	8/11/18 → 9/11/18

Bibliographical note

Green Open Access added to TU Delft Institutional Repository ‘You share, we take care!’ – Taverne project https://www.openaccess.nl/en/you-share-we-take-care

Otherwise as indicated in the copyright section: the publisher is the copyright holder of this work and the author uses the Dutch legislation to make this work public.

Access to Document

10.1007/978-3-030-31978-6_7

978-3-030-31978-6_7Final published version, 2.72 MB

Cite this

Oliehoek, F. A., Savani, R., Gallego, J., van der Pol, E., & Groß, R. (2019). Beyond Local Nash Equilibria for Adversarial Networks. In M. Atzmueller, & W. Duivesteijn (Eds.), Artificial Intelligence : 30th Benelux Conference, BNAIC 2018, Revised Selected Papers (pp. 73-89). (Communications in Computer and Information Science; Vol. 1021). Springer. https://doi.org/10.1007/978-3-030-31978-6_7

@inproceedings{6954eb472b2c4501872ad8e621e19846,

title = "Beyond Local Nash Equilibria for Adversarial Networks",

abstract = "Save for some special cases, current training methods for Generative Adversarial Networks (GANs) are at best guaranteed to converge to a {\textquoteleft}local Nash equilibrium{\textquoteright} (LNE). Such LNEs, however, can be arbitrarily far from an actual Nash equilibrium (NE), which implies that there are no guarantees on the quality of the found generator or classifier. This paper proposes to model GANs explicitly as finite games in mixed strategies, thereby ensuring that every LNE is an NE. We use the Parallel Nash Memory as a solution method, which is proven to monotonically converge to a resource-bounded Nash equilibrium. We empirically demonstrate that our method is less prone to typical GAN problems such as mode collapse and produces solutions that are less exploitable than those produced by GANs and MGANs.",

author = "Oliehoek, {Frans A.} and Rahul Savani and Jose Gallego and {van der Pol}, Elise and Roderich Gro{\ss}",

note = "Green Open Access added to TU Delft Institutional Repository {\textquoteleft}You share, we take care!{\textquoteright} – Taverne project https://www.openaccess.nl/en/you-share-we-take-care Otherwise as indicated in the copyright section: the publisher is the copyright holder of this work and the author uses the Dutch legislation to make this work public.; 30th Benelux Conference on Artificial Intelligence, BNAIC 2018 ; Conference date: 08-11-2018 Through 09-11-2018",

year = "2019",

doi = "10.1007/978-3-030-31978-6_7",

language = "English",

isbn = "978-3-030-31977-9",

series = "Communications in Computer and Information Science",

publisher = "Springer",

pages = "73--89",

editor = "Martin Atzmueller and Wouter Duivesteijn",

booktitle = "Artificial Intelligence",

}

Oliehoek, FA, Savani, R, Gallego, J, van der Pol, E & Groß, R 2019, Beyond Local Nash Equilibria for Adversarial Networks. in M Atzmueller & W Duivesteijn (eds), Artificial Intelligence : 30th Benelux Conference, BNAIC 2018, Revised Selected Papers. Communications in Computer and Information Science, vol. 1021, Springer, Cham, pp. 73-89, 30th Benelux Conference on Artificial Intelligence, BNAIC 2018, ‘s-Hertogenbosch, Netherlands, 8/11/18. https://doi.org/10.1007/978-3-030-31978-6_7

Beyond Local Nash Equilibria for Adversarial Networks. / Oliehoek, Frans A.; Savani, Rahul; Gallego, Jose et al.
Artificial Intelligence : 30th Benelux Conference, BNAIC 2018, Revised Selected Papers. ed. / Martin Atzmueller; Wouter Duivesteijn. Cham: Springer, 2019. p. 73-89 (Communications in Computer and Information Science; Vol. 1021).

Research output: Chapter in Book/Conference proceedings/Edited volume › Conference contribution › Scientific › peer-review

TY - GEN

T1 - Beyond Local Nash Equilibria for Adversarial Networks

AU - Oliehoek, Frans A.

AU - Savani, Rahul

AU - Gallego, Jose

AU - van der Pol, Elise

AU - Groß, Roderich

N1 - Green Open Access added to TU Delft Institutional Repository ‘You share, we take care!’ – Taverne project https://www.openaccess.nl/en/you-share-we-take-care Otherwise as indicated in the copyright section: the publisher is the copyright holder of this work and the author uses the Dutch legislation to make this work public.

PY - 2019

Y1 - 2019

N2 - Save for some special cases, current training methods for Generative Adversarial Networks (GANs) are at best guaranteed to converge to a ‘local Nash equilibrium’ (LNE). Such LNEs, however, can be arbitrarily far from an actual Nash equilibrium (NE), which implies that there are no guarantees on the quality of the found generator or classifier. This paper proposes to model GANs explicitly as finite games in mixed strategies, thereby ensuring that every LNE is an NE. We use the Parallel Nash Memory as a solution method, which is proven to monotonically converge to a resource-bounded Nash equilibrium. We empirically demonstrate that our method is less prone to typical GAN problems such as mode collapse and produces solutions that are less exploitable than those produced by GANs and MGANs.

AB - Save for some special cases, current training methods for Generative Adversarial Networks (GANs) are at best guaranteed to converge to a ‘local Nash equilibrium’ (LNE). Such LNEs, however, can be arbitrarily far from an actual Nash equilibrium (NE), which implies that there are no guarantees on the quality of the found generator or classifier. This paper proposes to model GANs explicitly as finite games in mixed strategies, thereby ensuring that every LNE is an NE. We use the Parallel Nash Memory as a solution method, which is proven to monotonically converge to a resource-bounded Nash equilibrium. We empirically demonstrate that our method is less prone to typical GAN problems such as mode collapse and produces solutions that are less exploitable than those produced by GANs and MGANs.

UR - http://www.scopus.com/inward/record.url?scp=85076089932&partnerID=8YFLogxK

U2 - 10.1007/978-3-030-31978-6_7

DO - 10.1007/978-3-030-31978-6_7

M3 - Conference contribution

AN - SCOPUS:85076089932

SN - 978-3-030-31977-9

T3 - Communications in Computer and Information Science

SP - 73

EP - 89

BT - Artificial Intelligence

A2 - Atzmueller, Martin

A2 - Duivesteijn, Wouter

PB - Springer

CY - Cham

T2 - 30th Benelux Conference on Artificial Intelligence, BNAIC 2018

Y2 - 8 November 2018 through 9 November 2018

ER -

Beyond Local Nash Equilibria for Adversarial Networks

Abstract

Publication series

Conference

Bibliographical note

Access to Document

Other files and links

Fingerprint

Cite this