ENGLISH | JAPANESE

1st International AI Competition on Werewolf Game @ANAC2019, IJCAI (pre-announcement)

2019/01/23

AIWolf Competition, international department will start.

This competition will be held in ANAC workshop at a top international conference on artificial intelligence IJCAI 2019 @ Macao.  This international competition is only for protocol division. Thus, if you just participate in this international convention, you do not need much English skills (you will need to register separately to participate in the international conference IJCAI itself).

Let’s try aiming for the first winner of the AIwolf competition. Currently, pre-registration survey is ongoing.

1. Outline

The Werewolf Game is a popular multiplayer discussion game, where each player needs to read the others’ intentions, while at the same time trying to persuade them to obtain cooperation. The AIWolf project aims for the development of AI agents for the Werewolf game. These agents must be able to exchange information, modeling the other’s knowledge and intention while hiding their own true win condition in a well defined communication channel.

The AIWolf project has conducted the AIWolf competition successfully in the CEDEC conference (Japanese national industry-academic conference on entertainment technologies). Each competition has gathered 100 to 200 participants. This gives us enough experience to run this kind of competition confidently. Our experiences running the Werewolf League in AAAI computer games workshop has been published at:

https://link.springer.com/chapter/10.1007/978-3-319-57969-6_8

This year, we want to extend the AIWolf competition to an international audience at ANAC. The AIWolf conference has two branches: the NLP branch, and the Protocol branch. For the ANAC competition, we recommend the Protocol branch, where the agents communicate with each other using a fixed protocol.

2. The Werewolf Competition (Protocol Division)

In the werewolf game, each agent is a member of a community (villager team and werewolf team). Each community has its own clear goal (eliminate all villagers/eliminate all werewolves), however the villager agents do not know the community affiliation of the other agents (the werewolves know the community affiliation of other agents). Additionally, each agent has a different role (such as villager, seer, medium, bodyguard, betrayer), which influences that agent’s available actions and their optimal strategy.

Because this mutual lack of information about each agent’s roles and affiliation, the agents in a werewolf game must negotiate to infer other agent’s reward tables, differentiate between cooperative and deceptive agents, and coordinate a voting and action strategy to reach the community’s goal.

This definition puts the werewolf game in the class of “negotiation” multi-agent games, such as Diplomacy. In Diplomacy, each agent (shown as a country) has its own clear goal, but their reward process is partially hidden to other agents. Each agent’s negotiation is focused in finding cooperative agents and constructing strategies for maximizing their final rewards.

Quick Summary of the Werewolf Game:

  • Agents are assigned to one of two asymmetric communities:
    • Team Werewolf: Minority of agents, know the identity of other agents in the same community. Optimal state: Remove all Villagers.
    • Team Villager: Majority of agents, do not know the identities of other agents in the same community. Optimal state: Remove all Werewolves.
  • Each agent is assigned a different (possibly repeated) role, which changes its possible actions and thus its optimal strategy: Villager, Seer, Medium, Bodyguard, Betrayer, etc. The role of an agent is known only to itself.
  • At each round, the following happens:
    • Negotiation Stage: Agents can exchange information through messages (in Protocol or Natural Language). Agents must build an optimal strategy to find cooperative agents and agree on a common strategy (voting and actions)
    • Voting Stage: Agents must vote to remove one agent from the game that round. The basic strategy is that villagers want to remove werewolves and vice-versa, but this strategy may be modified by cooperative/deceptive agents, roles, knowledge.
    • Action Stage: After the voting stage, some agents may take non-negotiating actions, such as protection, killing, investigation. These actions are determined by an agent’s role, and informed by information obtained during the negotiation stage of the round.

3. Preparations

The AIWolf Project has run its national competition for four years, so much of the necessary preparations have already been done, including Server Code, Client Code, Examples, Manuals, Webpage and Submission forms. Only minor updates are necessary as detailed below. The main page for the project is: http://aiwolf.org/en/, while the nightly version of the code for the contest are at: https://github.com/aiwolf/

3.1 Technical Aspects

The server and example client code are available at http://aiwolf.org/en/, under an MIT license. In past competitions, our rule was that all submissions should be in some form of open source license (MIT, GNU, CC), but we did not restrict which kind of open source license was used by the participant.

Participants initially submit client code following the specifications in the webpage above. The participants can download both sample agents and the server code, so that they can test their agent many times before the final submission. A participant’s submission will initially be evaluated by a contest between all participants, and some of the submissions of this contest will be selected for the final match. Those selected for the final match must also submit an explanation of their submission, to be presented at the ANAC Workshop.

For this contest, minor update to the public versions of the agent and server code are necessary, as well as a respective update to the English documentation. The dedicated league website is already available at http://aiwolf.org, and the webpage for contest submission is already available at http://contest.aiwolf.org/. These pages will be updates with details about the ANAC competition.

3.2 Communication Channels

The competitors will be in contact with the contest organization through the contest webpage (http://contest.aiwolf.org/), the contest github repository (https://github.com/aiwolf/), as well as the mailing list gm@aiwolf.org.

The CFP and contest announcement will be announced on relevant mailing lists of research societies (JSPS, Fuzzy, ALife, etc), as well as social media (Twitter, Facebook, etc). Additionally, in this year we have an accepted organized session on JSPS annual conference, where we can invite several people to send submissions to ANAC workshop.

3.3 ANAC Workshop

Based on our past AI-Wolf competitions in other academic events, the contest workshop at ANAC will be as follows: Each participant will be able to describe their agent, including their technologies and strategies. We will conduct one or more exhibition games (non-ranking) between the agents, followed by the announcement of the final result of the competition. Finally, we may announce/demo other technologies related to the AI-wolf project (such as a NLP protocol exhibition game)

3.4 Organization Schedule (Tentative)

  • Update of Protocol Division Server and Sample code: 2nd Week of January
  • Translation of Protocol Division Manual: 4th Week of January
  • ANAC competition website / FAQ / documentation: 2nd Week of January – 2nd Week of February
  • Announcement of Contest: Mid-February/Early March
  • Deadline for agent submission: July 15th
  • Agent Qualification: July 15-22
  • Final Match Notice: July 23rd
  • Deadline for submission of final agent: August 1st
  • Session/Awards Ceremony (@ ANAC): August 10-16
  • Preparation of competition Report: September
Share

2 responses to “1st International AI Competition on Werewolf Game @ANAC2019, IJCAI (pre-announcement)”

  1. hi,
    I’m trying to figure out the terms of the competition but i’m having hard time to find information in english…
    1. Should we implement a NLP mechanism for our agent? (sending & receiving messages to rivals) or should i just use the protocol?
    2. As part of our effort to set-up the AI-wolf framework: Does the presentation suit to the new 0.5.2 framework?

    thank you

    Reply

Leave a Reply

Your email address will not be published. Required fields are marked *