“AIWolf” is the AI agent that can play the game “Are you werewolf?” (“Werewolf” in short), which is popular among young people. It is one of the incomplete information games such as mahjong and poker, having information not disclosed to other players, in contrast to the complete information games such as chess and go where the strength of AI surpasses that of humans. Hidden information in the Werewolf game are the “role” determining the player’s special abilities and the “team” determined by their victory conditions. Furthermore, the Werewolf game also has the aspect of a “communication game” in which the game progresses only by conversation, making Werewolf game a very difficult game for AI.
The AIWolf Project is a project in which researchers from multiple universities and companies are promoting research on AIWolf. The project is proposing the Werewolf game as a standard problem for game AI in the future. Our laboratory is participating in this project, and is in charge of developing the AIWolf Platform and managing the AIWolf Competition.
AIWolf research is interesting
Research on AIWolf is a treasure trove of research themes necessary for the realization of “social intelligence” that will be required for AI in the near future. It goes beyond simply creating strong game AI, to AI technology that deceives or persuades people or other AI, or AI technology that takes actions that entertain people.
Here, we will introduce research aimed at strong AIWolf agent using deep learning. In the Werewolf game, how to estimate the roles of other players determines the outcome of the game. We converted past game records called “Game Logs” into features and learned them with a deep neural network. The estimation accuracy of the role estimator obtained by learning achieved about 90% in most roles. Our agent using this role estimator achieved a good result of 2nd place in the 2nd International AIWolf Competition. In particular, the fact that the winning rate in the role of “villager” who does not have special abilities became the highest among the participating teams shows the power of this role estimator.