3rd International Aiwolf Contest
3rd International AIWolf Contest
Games are a great platform to develop Artificial Intelligence research. The social game Werewolf (aka Mafia) is a particularly difficult game for which no AI that can play at a strong level has been found yet. The difficulties of this game include the social factor, the communication based gameplay, and the necessity of modeling the internal belief state of all other agents
To promote research in Werewolf AI and move towards the solution to these challenges, we are holding the Third International Werewolf AI Competition (AIWolf Contest), in tandem with the Automated Negotiating Agents Competition (ANAC2021). The results will be announced at IJCAI 2021.
The goal of this competition is to develop an AI agent for the Werewolf game, that can perform reliably well against other AI agents. The competition has two sub divisions: In the Protocol Division, the agents communicate using a fixed artificial protocol. In the Natural Language division, the agents communicate in natural language texts (Japanese or English).
Protocol Division Deadline (updated 6/24)
- Participant Registration: July 5th (Mon) Registration site
- Agent Submission: July 19th (Mon)
- Finalist Announcement: Early August
- Finalist Agent Submission: August 13th (Fri)
- Final Result Announcement: IJCAI, August 21th to 26th.
NLP Division Deadline (Updated 6/24)
- Registration and Self-play log submission: July 30th (Fri)
- NLP Finalist Games: August 1st Week
- 2021/06/24: Deadline extended.
- 2021/04/17: Updated information of the Natural Language Division.
- 2021/04/02: Regulation update to version 1.2.2
- 2021/03/23: 3rd International Competition website is up!
Protocol Division Information
- The Call for Participation (v 1.0) describes the overall competition.
- The Competition Regulation (v 1.2.3) describes the game rules, the contest rules, and the registration and submission procedure.
- The AIWolf Protocol (v. 3.6) describes the artificial language that agents can use to talk to each other.
- Register your team and agent on the Registration Site. You can also play test games on the server to check your agent for bugs. Highly recommended!
- The competition will use version 0.6.2 of the AIWolf server.
- You can write your agent in Java, Python or .NET
- Java Resources
- Python Resources
- AIWolfPy Library
- AIWolf Python Programming Tutorial (Video)
- Python modules installed in the competition server. Contact us if you need other modules.
- .NET resources
- You can also download agents from past competitions: 1st Intl. Competition, 2nd Intl. Competition.
- You can also download logs from games with the agents from the last competition: 2nd Int. Competition Game Logs
Natural Language Division Information
In this division, participants must build an artificial intelligence agent that can play the werewolf game as humans do, using natural language. Participants in this division will be evaluated by a panel of judges, who will grade the subjective quality of the dialog generated by the agent. Agents must communicate in Japanese or English.
How to participate:
A team should send a mail to aiwolf [at] kanolab.net (replace at by @), describing your team name, a contact e-mail address, names and affiliations of its members (please mark a contact person when a team consists of multiple members). Registration is free.
- Agent Testing
A participant is required to implement an AI werewolf agent that connects to our AIWolf server at the specified time. Before that, we will provide an AIWolf server where participants can try connecting with dummy agents to check their system’s behavior. Make sure that your system works correctly on the testing system before the actual competition run.
- Preliminary Run
Participants should run five agents connected to our server, running 50 games. Submit your game logs to the organisers via e-mail. If there are too many participants for the final run, organisers will select final run teams from these logs. In either case, these logs will be used in the final evaluation.
- Final Run
Each participant team is required to connect its agent to our server at a specified time to play games with other participants. If your team cannot remotely attend the final run, you should prepare your system so that the organisers can somehow trigger your remote system to connect our server and run, for example, via a web browser.
- System Evaluation
Participants should submit a system design description document to the organizers. This document and logs of the games might be used for research purposes and included and published in an overview paper without any further permission. Participants are encouraged to submit a paper to the ANAC workshop. Reviewers will perform subjective evaluations on the game logs, using following criteria:
- A Natural utterance expressions
- B Contextually natural conversation
- C Coherent (not contradictory) conversation
- D Coherent game actions (vote, attack, divine) with conversation contents
- E Diverse utterance expressions, including coherent characterization Please note that vague utterances that could be used regardless of context are not always natural in the werewolf game.
- Implementation of Natural Language Division Agents: Agents must communicate in English or Japanese. We might translate English-Japanese by machine translation internally when the number of participants for either of the languages is small. Participants are required to implement an AI werewolf agent that connects to our AIWolf server via the network. That is, there is no limitation on the client side regarding hardware or software other than the network I/O specifications.
- Please refer to our sample agent code with runtime environment (java), distributed from our Github repository:
which also describes how to connect to our game server, etc.
The Mafiascum Dataset provides is a collection of over 700 games of Mafia:
Links and Contact Information