While humans play the werewolf game using natural language, it is very difficult to design effective agents that can use Natural Language easily to play werewolf. Additionally, unlike day-to-day dialogue where the focus is to describe the agent’s environment, the werewolf game uses very specific language constructs that makes it difficult to just apply existing Natural Language Processing techniques.
With these reasons in mind, we have developed a protocol that can be used to mediate games between humans and agents, and games among agents themselves. This protocol describes how an agent can express its internal thought model to other agents, and how they can infer information from their communications.
In this page, we define the grammatical structure and necessary keywords for the “AIWolf Protocol”, which can express the necessary conversation for a Werewolf game.
- AIWolf Protocol ver3.6 (Latest Version)
- Now it is possible to leave the “subject” and “target” of sentences undefined by using the ANY keyword.
- Added expressions to report the past target of a vote or attack (VOTED, ATTACKED)
- Modified the structure of REQUEST sentences.
- Added 7 new keywords. (INQUIRE, BECAUSE, DAY, NOT, AND, OR, XOR)
- AIWolf Protocol ver2.01 (2017, in Japanese)
- Added the REQUEST keyword. Extended the grammatical structure to accommodate for composite request sentences.
- As the protocol becomes more complex, we have added examples of common sentences in Werewolf and how to express them in the protocol (In Japanese).
- AIWolf Protocol ver1.11（2014-01-27）
- AIWolf Protocol Ver1.00（2014-01-27）