Facebook's A.I. Masters the Card Game Hanabi

Facebook has made a major advancement with its artificial intelligence software by teaching it to master Hanabi, a Solitaire-like card game that requires players to work together.

Digital bots playing the game learned to assist each other to achieve higher scores than previous A.I. systems, the social networking giant said on Friday.

Although creating A.I. that dominates a card game like Hanabi seems minor, it’s actually a big deal for researchers. The same A.I. that helps digital bots learn Hanabi, in theory, could be used to help virtual assistants interact more intuitively with people in the real-world.

In recent years, some of the biggest advancements in A.I. have come from researchers creating software that wins complicated games. For example, in 2016 Google’s DeepMind’s A.I. system defeated leading human players in the ancient board game Go.

Now, Hanabi is being touted as a capable testbed for A.I. because it requires teamwork and strategy. It presents newer challenges for A.I. researchers who are looking to build more sophisticated A.I. systems that can do more than defeat opponents.

“One of the reasons we’re moving to these cooperative games is that I think we’re kind of at the point where there’s no games left at least in terms of competitive games,” said Adam Lerer, a Facebook researcher who worked on the paper.

In the game of Hanabi, teams of two to five players are given random cards of different colors and numbers that represent points. The goal for teams is to lay the cards on a table, grouped by color, in the correct numerical order.

The problem, however, is that players cannot see their own cards while their teammates can. A player can give hints to another, like making a remark about a certain color, that would tip the other off to do something like playing or discarding a card. The dilemma is that the player must deduce what their teammate’s clue means.

The challenge for Facebook’s A.I. researchers was to give their Hanabi bots a way to understand the hints of their teammate bots based on the limited information they have about their own cards.

To help the Hanabi bots, Facebook used what’s known as a Monte Carlo “search” technology technique to help them evaluate their possible moves. The Monte Carlo search technique was previously used by DeepMind researchers to help their powerful AlphaZero- A.I. software navigate games like Chess and Go. This summer, Facebook and Carnegie Mellon University used the same search technique to create their poker-beating Pluribus A.I. software.

For Hanabi, Facebook researchers created a variant of Monte Carlo that lets multiple Hanabi bots evaluate multiple playing options while sharing information with each other. The combination of both reinforcement learning and the so-called multi-search technique helped Facebook’s bots learn to play Hanabi on their own and in tandem with one another.

Reinforcement learning was good at helping the Hanabi bots come up with a “decent strategy” to master the game, while the search technique helped the bots refine that approach, he explained.

It’s possible that some of the techniques used to create the Hanabi bots could one day serve to advance self-driving cars. For instance, cars embedded with variations of the technology could more safely navigate roads by taking in account the actions of other drivers.

“So if you think about self-driving cars, you have to know how to communicate your actions,” Lerer said. “There’s a ‘theory-of-mind element where you have to reason about the world based on what other people are doing.”

Facebook researchers will present a findings about their research paper next week during the annual Neural Information Processing Systems conference in Vancouver. It is to be published in 2020 by the Association for the Advancement of Artificial Intelligence.

Facebook’s A.I. Masters the Card Game Hanabi

More must-read stories from Fortune: