Emergent password signalling in the game of Werewolf
Emergent password signalling in the game of Werewolf
Emergent communication can lead to more efficient problem-solving heuristics and more domain specificity. It can perform better than a handcrafted communication protocol, potentially directing autonomous agents towards unforeseen yet effective solutions. Previous research has investigated a social deduction game, called Werewolf, where two groups of autonomous agents, villagers and werewolves, interact in an environment named RLupus. We study the impact of allowing the agents to communicate through multiple rounds and evaluate their language and performance against the baseline environment. We show that agents develop a highly successful heuristic using a single word vocabulary. They create an approach using passwords, allowing them to determine which agents are werewolves, which is the winning condition. We explore the possible reasons behind this strategy, with further experimental analysis showing that our approach speeds up the convergence of the agents towards a common communication strategy.
emergent communication, Reinforcement Learning, artificial intelligence (AI)
Lipinski, Olaf
88709b3f-c356-45c7-8520-cb49d7b07960
Sobey, Adam
e850606f-aa79-4c99-8682-2cfffda3cd28
Cerutti, Federico
65aba5ac-fb31-47c4-a585-80b5c4e1bcac
Norman, Timothy
663e522f-807c-4569-9201-dc141c8eb50d
29 April 2022
Lipinski, Olaf
88709b3f-c356-45c7-8520-cb49d7b07960
Sobey, Adam
e850606f-aa79-4c99-8682-2cfffda3cd28
Cerutti, Federico
65aba5ac-fb31-47c4-a585-80b5c4e1bcac
Norman, Timothy
663e522f-807c-4569-9201-dc141c8eb50d
Lipinski, Olaf, Sobey, Adam, Cerutti, Federico and Norman, Timothy
(2022)
Emergent password signalling in the game of Werewolf.
Emergent Communication Workshop at ICLR 2022.
29 Apr 2022.
Record type:
Conference or Workshop Item
(Paper)
Abstract
Emergent communication can lead to more efficient problem-solving heuristics and more domain specificity. It can perform better than a handcrafted communication protocol, potentially directing autonomous agents towards unforeseen yet effective solutions. Previous research has investigated a social deduction game, called Werewolf, where two groups of autonomous agents, villagers and werewolves, interact in an environment named RLupus. We study the impact of allowing the agents to communicate through multiple rounds and evaluate their language and performance against the baseline environment. We show that agents develop a highly successful heuristic using a single word vocabulary. They create an approach using passwords, allowing them to determine which agents are werewolves, which is the winning condition. We explore the possible reasons behind this strategy, with further experimental analysis showing that our approach speeds up the convergence of the agents towards a common communication strategy.
Text
EmeCom_at_ICLR_2022___Werewolf
- Version of Record
More information
Published date: 29 April 2022
Venue - Dates:
Emergent Communication Workshop at ICLR 2022, 2022-04-29 - 2022-04-29
Keywords:
emergent communication, Reinforcement Learning, artificial intelligence (AI)
Identifiers
Local EPrints ID: 457013
URI: http://eprints.soton.ac.uk/id/eprint/457013
PURE UUID: f44ea883-96d3-4aa1-8a04-0d87e9cb9820
Catalogue record
Date deposited: 19 May 2022 16:44
Last modified: 17 Mar 2024 04:03
Export record
Contributors
Author:
Olaf Lipinski
Author:
Federico Cerutti
Download statistics
Downloads from ePrints over the past year. Other digital versions may also be available to download e.g. from the publisher's website.
View more statistics