Playing hanabi near-optimally

Abstract

This paper describes a study on the game of Hanabi, a multi-player cooperative card game in which a player sees the cards of the other players but not his own cards. Previous work using the hat principle reached near-optimal results for 5 players and 4 cards per player: the perfect score was reached 75% of times on average. In the current work, we develop Hannibal, a set of players, aiming at obtaining near-optimal results as well. Our best players use the hat principle and a depth-one search algorithm. For 5 players and 4 cards per player, the perfect score was reached 92% of times on average. In addition, by relaxing a debatable rule of Hanabi, we generalized the near-optimal results to other numbers of players and cards per player: the perfect score was reached 90% of times on average. Furthermore, for 2 players, the hat principle is useless, and we used a confidence player obtaining high quality results as well. Overall, this study shows that the game of Hanabi can be played near-optimally by the computer.