Hanabi

Hanabi is a cooperative card game for two to five players, in which the players are dealt cards in five different colors with ranks from 1 to 5 representing fireworks. Unlike in most card games, players hold their cards facing away from themselves, i.e. they can not see their own cards, but they can see every other player’s cards. The goal of the game consists in playing the cards starting with the 1s in each color, continuing in ascending order up to the 5s. The game restricts what the players can say, only allowing them to give very specific types of hints to other players. For example, a player may tell another player where all of their 1s are, or all of their red cards. While communication in the game is very restrictive, human players tend to use a similar approach to how they communicate in other settings, following Grice’s maxims of communication. Because of its simple rules, yet challenging game play requiring very efficient use of a very limited communication channel, Hanabi has garnered a lot of interest from the AI research community, with DeepMind calling it a new frontier for AI research.

As part of Dr. Markus Eger’s dissertation research, he developed AI agents that used Grice’s maxims of communication to expose intentional behavior in a manner that is understood by human players. The resulting agents then played with over 200 participants in a controlled experiment, with the newly developed agents playing very well with human players, as reported in this paper. Additionally, the source code for the Hanabi agents, as well as game logs from participants that consented to it were made publicly available on GitHub.

We have been continuing research on these agents, and worked on improving their communication capabilities. In a current project, we have been incorporating the time human player’s thought about their actions into the agent’s decision making process, which allowed us to improve how the agent interprets information it receives. A publication with the results of this experiment is currently in preparation.