In Yokohama on August 31 and September 1, an artificial intelligence tournament invites contestants to develop manipulative impostor bots.
From our correspondent in Tokyo
After the abstract strategy of Go, bluffing in Texas Hold’em Poker and even the visual interpretation of a mock Pictionary, the artificial intelligence of Japanese chatbots is tackling group psychology with the “Artificial Intelligence Werewolf” (AIWolf) project.
Werewolf, a thematic adaptation of the original Mafia game invented by a Russian psychology professor in 1987, is a party game that, once the cards are dealt, is based entirely on the players’ discussion—and lies. The game can be played either in a physical room together or online. The players are divided into two camps: on one side, the villagers cooperate to unmask and kill the werewolves; on the other, the werewolves conspire to kill the villagers at night, while pretending to be “innocent” villagers during the day. It goes without saying that the werewolves must bluff and lie in order to survive.
Werewolf stimulates artificial intelligence
This confrontational role-playing game presents several challenges to artificial intelligence. First, the bots must communicate in natural, even persuasive, language that goes well beyond the Turing test. Second, Werewolf is an incomplete information game, like poker, where players must make decisions based on only partially revealed information. Furthermore, the players’ knowledge is “asymmetrical” in the sense that the werewolves know more than the villagers, who can only guess who are the werewolves among them. As such, the AI research involved ranges from machine learning to logic programming, language processing and cognitive science.
The AIWolf project, launched in 2013 by a group of Japanese researchers in artificial intelligence, offers an open source platform for developers to create virtual agents that play the game online while communicating naturally (although not necessarily truthfully) and (re)acting according to human intuition and reasoning—chatbots pretending to be werewolves pretending to be villagers…
AIWolf is supported primarily by the game publisher Oink Games, the young Tokyo company dedicated to the Werewolf game Jinraw, and a research professor at the University of Electro-Communications, co-author of a book on artificial intelligence applied to Werewolf, Kosuke Shinoda.
Since 2014, in parallel to collectively developing the virtual agents, the group organizes the AI Werewolf Mini Games during the annual Game AI Tournament at the University of Electro-Communications, and every summer the big AI Werewolf Contest at the Computer Entertainment Developers Conference (CEDEC), in addition to presentations, seminars and workshops on artificial intelligence applied to Werewolf throughout the year.
Mind-reading algorithms
The 3rd AI Werewolf Contest will be held in Yokohama at CEDEC 2017 on August 31 and September 1, with a grand prize of ¥100,000 ($900). The game contest will play out in a village of 5 people, including 2 villagers, 1 werewolf, 1 seer and 1 madman, with a 5-second response time (no punctuation or emoticons allowed). The agents can be created in Java, .NET or Python languages. The complete contest rules and conditions (in Japanese, last updated on June 17, 2017), as well as program libraries and two natural language agents developed for the Mini Games in March, can be downloaded from the AIWolf website. Registration for teams is free and open until July 14. Preliminary rounds will take place at the end of July.
Given the particular challenges of the Werewolf game and the individual personalities of every human player, we’re more than a little anxious to meet the next agents trained by deep learning—mind-reading algorithms, capable of deceiving us with their lies while calling our bluff, rallying the group to eliminate us. Manipulative, lie-detecting, lynch-mobbing impostor chatbots… the next frontier in artificial intelligence?
More about AIWolf (in English and Japanese)