- The impact of artificial intelligence on the workforce is a hotly debated subject. In the world of gaming, however, AI is already replacing humans.
- Facebook AI researchers say they built a bot that can exceed human performance in Hanabi, a cooperative game that requires players to interpret the actions and intentions of other players.
- The achievement signals a jump in the capability of the technology to replicate the cognitive skills of humans, as opposed to the method of training the models through rewards and punishments that was used for other models.
- Click here for more BI Prime stories.
Last month, Lee Se-dol — one of the top South Korean players in Go, a complex Chinese board game that dates back 2,500 years — announced he would retire because it became impossible to beat AI-powered bots. This comes after Se-dol lost a series of matches in 2016 to a bot developed by Google-owned AlphaGo.
Now, Hanabi may be next. Facebook researchers have built a bot they say can outperform the most skilled players at the cooperative card game that requires individuals to play off the hands of other participants because they are unable to view their own cards.
While other models have exceeded at games like Go, chess, and poker, researchers were able to scale those projects through a technique known as "reinforcement learning," a process that uses a series of rewards and punishments to train behavior — similar to how one might approach raising a puppy.
Project leads at the social media giant's research lab say their latest bot signals a significant improvement because it can replicate cognitive skills to put itself in the perspective of other players in the game, a trait known as "theory of mind."
"These techniques that we are developing that can cope with things like theory of mind will play an important part as we extend to other domains like natural language processing," AI researcher Noam Brown told Business Insider.
Employing AI to cooperate, instead of compete, against humans
Facebook AI Research is a group of about 300 employees that is wholly focused on academic studies into the technology. While some of the projects might dovetail off of Facebook's AI goals, Brown and fellow researcher Adam Lerer say they independently choose what problems to tackle.
The group aims to answer "long-term research questions that we think of as barriers to the advancement of artificial intelligence," said Lerer.
"You can't always predict what is going to be the next big thing, so you have to give the freedom to the researchers … even if there's not a five-year plan for turning this into a product" that can be monetized, Brown added.
Hanabi posed a unique challenge because unlike chess, for example, the bot is not going strictly head-to-head against a human. The nature of the game requires players to cooperate with other participants and interpret their intentions based on gameplay to win.
A player is unable to view their hand — just the hands of the other participants — and only narrow hints are provided throughout gameplay. So one player may seek to gain an advantage by suggesting one hidden card is a certain color because he or she wants that color played next.
Hanabi was proposed as a key challenge for AI by researchers at Google's DeepMind for that reason.
The model had to reason "about the intent and the beliefs that other people have that lead them to take the actions that they do." Lerer said. The team had to build a "conversational agent that can reason about what a person is saying without having to explicitly tell them everything."
The project is so promising because it signals that AI could eventually be able to cooperate alongside humans to tackle complex challenge, as opposed to just replacing them in tasks like monitoring shelves in Walmart for restocking.