The AIWolf Project aims to create an Artificial Intelligence agent that is capable to play the Werewolf game in place of a human player.
Werewolf (Also called Mafia) is an assymetric information game.
In its simplest formulation, two teams (Werewolf Team and Villager Team) try to eliminate all players of the opposing team. This is done through a repeated voting process, in which all players discuss and eventually vote on a player to be eliminated from the game.
The twist is that the Villager team has more members, who are not aware of each other's identities, while the Werewolf team has a small number of members, who know each other's identities. This makes the werewolf a contest of an informed minority against an uninformed majority.
In terms of strategy, if the Villager team is able to identify the werewolf players, they can easily use their larger numbers to vote and eliminate the werewolfs. On the other hand, the werewolf team can use their information about who is and who is not a werewolf to manipulate the players into voting on players of the Villager team. In this sense, from the point of view of a Villager player, the game is one of investigation, while from the point of view of a Werewolf player, the game is one of deceit.
Finally, in adition of voting and elimination, the Werewolf game usually includes Roles and special actions that can modify the game dynamic. These abilities occur during a "night phase", which occurs after each voting stage.
Some common dynamics are: The werewolf team is able to Attack and Kill one player during the night phase; There is a "seer" (or "cop") role in the villager team who can discover the role of another player during the night phase; There is a "bodyguard" (or "doctor") role in the villager team who can protect one player from the werewolf attack during the night phase; etc.
There are many variations and special roles in werewolf games. In the AIWolf project, the game server implements the following roles:
- Villager: regular villager role, no special abilities.
- Seer: at each night phase, has the ability to investigate one other player and learn whether they are a werewolf or a villager.
- Medium: at each night phase, has the ability to investigate one player who was eliminated and learn whether they are a werewolf or a villager.
- Bodyguard: at each night phase, has the ability to protect one player (including themselves) against werewolf attacks.
- Werewolf: regular werewolf role. During the night phase, all the werewolves can collectively choose one player to eliminate.
- Possessed: a regular villager, but aligned with the werewolf team. The seer's investigation will not identify a possessed as a werewolf.
Following the 4th AIWolf Competition rules, there are two format of games: 5 player game, or 15 player game. The role distribution for each format is as follows:
| Role | 5 Players | 15 Players |
|---|---|---|
| Villager | 2 | 8 |
| Seer | 1 | 1 |
| Possessed | 1 | 1 |
| Werewolf | 1 | 3 |
| Medium | - | 1 |
| Bodyguard | - | 1 |
The following fluxogram describes the flow of a Werewolf game in the competition platform:
The official distribution of the AIWolf framework and the game server is in Java. Therefore, the primary way of developing agents for the competition is by using java files offered on the official website. Basic instructions in English on which classes to inherit from are provided in the official website.
However, since the AIWolf clients connect to the server via sockets, in principle it should be possible to develop an agent in any language, as long as the proper protocol is followed. For the purposes of the competitions, as the organizers need to read and execute the code of the agents, only a few languages are officially supported. Python is one of them.
The first Python implementation for AIWolf can be found at Kei Harada's repository. The current repository has the same code with minor changes.
Before running your agent, you need to set up a server.
First download the latest server files from the website (aiwolf-ver0.5.*).
The first and simplest way to execute the server is using the GUI mode. Simply navigate to the unzipped folder and run
./StartServer.sh
This will open the GUI interface for the server. From there, you can select the number of players, press the "Connect" button, and the server will be ready for agents to connect.
In addition to connecting your own agents, you can also run and connect the sample java agents provided by the competition by using the StartGUIClinet.sh script.
Using the GUI mode you can not automatically run a certain number of games, which can be a problem for testing or training agents.
To get around this problem, you should use the terminal mode. First you should edit the AutoStarter.ini file and configure the terminal mode parameters.
lib=./
log=./log/ #folder to save logs
port=10000 #port to connect
game=10 #number of games to run
view=false #show the game GUI
setting=./SampleSetting.cfg #game related settings
agent=5 #number of agents to play
Sample1,org.aiwolf.sample.player.SampleRoleAssignPlayer #list of java agents to connect
Sample2,org.aiwolf.sample.player.SampleRoleAssignPlayer #jar files must be on the same folder
Sample3,org.aiwolf.sample.player.SampleRoleAssignPlayer
Sample4,org.aiwolf.sample.player.SampleRoleAssignPlayer
If you want to connect your python-based agent, simply leave less java-based clients in the agent list than the total number of agents specified (alternatively, leave the agent list empty and connect only your clients). The server is going to wait for connections until the specified number of agents is met. After configuring the AutoStarter.ini file, you can execute the server in terminal mode by running
./AutoStarter.sh
To execute the python script, you need a few libraries including numpy, pandas, etc. To check which libraries and which versions you need, check the requirements.txt file in the root folder of this repository. You can install all of them at once by running:
pip install -r requirements
It is highly recommended to setup a virtual environment before running this command, to assure there will be no conflicits with libraries already installed in your system. You can check how to install virtualenv and how to create a new environment here.
Once the server is set up and you have the required libraries, you can connect the sample python agent in the current repository by executing
./aiwoof.py -h localhost -p 10000
You can also request a role to the server by passing -r [ROLE] as an argument (roles can be VILLAGER, SEER, MEDIUM, BODYGUARD, POSSESSED, WEREWOLF).