Evolving HyperNetworks for Game-Playing Agents

Christain Carvelli; Djordje Grbic; Sebastian Risi

doi:10.1145/3377929.3389961

Evolving HyperNetworks for Game-Playing Agents

Christain Carvelli, Djordje Grbic, Sebastian Risi

Robotics, Evolution, and Art Lab

modl.ai

Publikation: Konference artikel i Proceeding eller bog/rapport kapitel › Konferencebidrag i proceedings › Forskning › peer review

Abstract

This work investigates the evolution of indirectly-encoded neural networks through a hypernetwork approach. We find that for some Atari games, a hypernetwork with over 14 times fewer parameters, can compete or even outperform directly-encoded policy networks. While hypernetworks perform worse than directly encoded networks in the game Frostbite, in the game Gravitar, the approach reaches a higher score than any other evolutionary method and outperforms complicated deep reinforcement learning setups such as Rainbow.

Originalsprog	Engelsk
Titel	GECCO '20: Proceedings of the 2020 Genetic and Evolutionary Computation Conference Companion
Antal sider	2
Forlag	Association for Computing Machinery
Publikationsdato	8 jul. 2020
Sider	71-72
ISBN (Trykt)	9781450371278
DOI	https://doi.org/10.1145/3377929.3389961
Status	Udgivet - 8 jul. 2020
Begivenhed	GECCO 2020: The Genetic and Evolutionary Computation Conference - online, Cancun, Mexico Varighed: 8 jul. 2020 → 12 jul. 2020 https://gecco-2020.sigevo.org/index.html/HomePage

Konference

Konference	GECCO 2020
Lokation	online
Land/Område	Mexico
By	Cancun
Periode	08/07/2020 → 12/07/2020
Internetadresse	https://gecco-2020.sigevo.org/index.html/HomePage

Emneord

Machine learning
Evolutionary algorithms
Artificial intelligence
Hypernetworks
Game-playing agents

Adgang til dokumentet

10.1145/3377929.3389961

carvelli_grbic_risi_2020Accepteret manuskript, 608 KB

Citationsformater

@inproceedings{03c4604397844720a14478e3a6855258,

title = "Evolving HyperNetworks for Game-Playing Agents",

abstract = "This work investigates the evolution of indirectly-encoded neural networks through a hypernetwork approach. We find that for some Atari games, a hypernetwork with over 14 times fewer parameters, can compete or even outperform directly-encoded policy networks. While hypernetworks perform worse than directly encoded networks in the game Frostbite, in the game Gravitar, the approach reaches a higher score than any other evolutionary method and outperforms complicated deep reinforcement learning setups such as Rainbow.",

keywords = "Machine learning, Evolutionary algorithms, Artificial intelligence, Hypernetworks, Game-playing agents, Machine learning, Evolutionary algorithms, Artificial intelligence, Hypernetworks, Game-playing agents",

author = "Christain Carvelli and Djordje Grbic and Sebastian Risi",

year = "2020",

month = jul,

day = "8",

doi = "10.1145/3377929.3389961",

language = "English",

isbn = "9781450371278",

pages = "71--72",

booktitle = "GECCO '20: Proceedings of the 2020 Genetic and Evolutionary Computation Conference Companion",

publisher = "Association for Computing Machinery",

address = "United States",

note = "GECCO 2020 : The Genetic and Evolutionary Computation Conference, GECCO2020 ; Conference date: 08-07-2020 Through 12-07-2020",

url = "https://gecco-2020.sigevo.org/index.html/HomePage",

}

Evolving HyperNetworks for Game-Playing Agents. / Carvelli, Christain; Grbic, Djordje ; Risi, Sebastian.
GECCO '20: Proceedings of the 2020 Genetic and Evolutionary Computation Conference Companion. Association for Computing Machinery, 2020. s. 71-72.

Publikation: Konference artikel i Proceeding eller bog/rapport kapitel › Konferencebidrag i proceedings › Forskning › peer review

TY - GEN

T1 - Evolving HyperNetworks for Game-Playing Agents

AU - Carvelli, Christain

AU - Grbic, Djordje

AU - Risi, Sebastian

PY - 2020/7/8

Y1 - 2020/7/8

N2 - This work investigates the evolution of indirectly-encoded neural networks through a hypernetwork approach. We find that for some Atari games, a hypernetwork with over 14 times fewer parameters, can compete or even outperform directly-encoded policy networks. While hypernetworks perform worse than directly encoded networks in the game Frostbite, in the game Gravitar, the approach reaches a higher score than any other evolutionary method and outperforms complicated deep reinforcement learning setups such as Rainbow.

AB - This work investigates the evolution of indirectly-encoded neural networks through a hypernetwork approach. We find that for some Atari games, a hypernetwork with over 14 times fewer parameters, can compete or even outperform directly-encoded policy networks. While hypernetworks perform worse than directly encoded networks in the game Frostbite, in the game Gravitar, the approach reaches a higher score than any other evolutionary method and outperforms complicated deep reinforcement learning setups such as Rainbow.

KW - Machine learning

KW - Evolutionary algorithms

KW - Artificial intelligence

KW - Hypernetworks

KW - Game-playing agents

KW - Machine learning

KW - Evolutionary algorithms

KW - Artificial intelligence

KW - Hypernetworks

KW - Game-playing agents

U2 - 10.1145/3377929.3389961

DO - 10.1145/3377929.3389961

M3 - Article in proceedings

SN - 9781450371278

SP - 71

EP - 72

BT - GECCO '20: Proceedings of the 2020 Genetic and Evolutionary Computation Conference Companion

PB - Association for Computing Machinery

T2 - GECCO 2020

Y2 - 8 July 2020 through 12 July 2020

ER -

Evolving HyperNetworks for Game-Playing Agents

Abstract

Konference

Emneord

Adgang til dokumentet

Fingeraftryk

Citationsformater