If you click on a link and make a purchase we may receive a small commission. Learn more.

Facebook has created an AI that can play Hanabi

Making fireworks with a robot.

Artificial Intelligence researchers at Facebook have created an AI bot that can play Hanabi better than other human players.

Despite not being the first AI developed specifically to play Hanabi this particular bot has outperformed all previous examples to master the game’s co-operative mechanics. This achievement in artificial intelligence research is especially commendable because previously created AI had struggled with situations wherein collaborating with others was important.

Fellow AI researchers at Deepmind and Google Brain were interested in Hanabi’s potential for experimenting with artificial intelligence, because playing the game means having to interpret other players’ communications in order to fill-in the missing gaps in your own body of information.

In Hanabi players work together to play different sets of cards in ascending numerical order, with each player able to see everyone else’s hand but their own. This is where the need for communication comes in as players take turns to give each other hints about what cards they hold. Developing an AI capable of comprehending subtle player-cues and actions was an entirely new challenge for Facebook’s researchers, but they managed to find a solution by implementing a real-time search method.

This search method was previously used by the bot Pluribus to successfully defeat a group of professional Texas Hold’em poker players. By rifling through the many different possible ‘world states’ of the game (or outcomes) the AI can use player actions to inform which decision will lead to the best possible path to victory.

When developing previous AI board game champions such as the AlphaGo AI that beat one of the world’s best Go players, this search method was used to perfect decision-making in games wherein all information was available to all players (as in chess and backgammon). But this new AI had to interpret other players’ actions to deduce their intentions and ultimately understand how best to win, a behaviour known as the theory of the mind.

With this newest AI Facebook hopes to further develop artificial intelligences capable of interpreting non-verbal cues and visualising human’s intentions, so that they can move further towards creating ‘helpful cooperative AI agents in the real world’.

But how many of us really just want an extra player for a few rounds of Hanabi?


Alex Meehan avatar

Alex Meehan

Staff Writer

Alex’s journey to Dicebreaker began with writing insightful video game coverage for outlets such as Kotaku, Waypoint and PC Gamer. Her unique approach to analysing pop culture and knack for witty storytelling finally secured her a forever home producing news, features and reviews with the Dicebreaker team. She’s also obsessed with playing Vampire: The Masquerade, and won’t stop talking about it.