Facebook is no stranger to artificial intelligence. The social media company uses a form of A.I., known as deep learning, to build its facial recognition software, and has developed a yet-to-be-released A.I. system that will describe photos to blind Facebook users. It makes a lot of sense that the company would try to stay on the cutting edge of both technology and social media by tinkering with A.I. systems. What’s surprising, however, is that their latest A.I. experiment is geared not towards keeping up with the latest social media trends, but towards winning a 2500-year-old board game.
According to WIRED, Facebook is trying to build an A.I. system that will be able to beat the world’s best Go players. Over the last few decades, computers have defeated the world’s best human players at checkers, chess, Scrabble, and even Jeopardy. In 1996, IBM’s Deep Blue supercomputer beat world champion chess player Gary Kasparov in one of the most famous chess games ever played. But no computer has ever been created that can beat humans at the ancient Chinese board game Go—and not for lack of trying.
Go is a deceptively simple game. Players can place their pieces on any intersection of two lines on the 19-by-19 grid, using lines of contiguous pieces to delineate territories, or to capture their opponent’s pieces. Whoever captures the most space and the most pieces by the end of the game is the winner. Unlike chess, pieces don’t move around the board in complex patterns—in fact, they don’t move at all unless captured. But the conceptual simplicity of the game is exactly what makes Go so tough: since players can place pieces anywhere on the grid, the range of different strategies and possible moves is immense.
WIRED explains that, traditionally, computers defeat humans at board games by “analyzing the many possible outcomes of every possible move.” Deep Blue beat Kasparov because it could automatically analyze more moves than he could, in a much smaller amount of time. But the range of possible moves on a Go board is not only too great for most computers to analyze while actively playing, but the exact rules behind what makes a good move can be difficult to articulate.
“We’re pretty sure the best [human] players end up looking at visual patterns, looking at the visuals of the board to help them understand what are good and bad configurations in an intuitive way,” Facebook CTO Mike “Schrep” Schroepfer told WIRED.
So, Facebook is using deep learning to develop a new approach to mastering Go. The company wants to build a system that will incorporate the “intuitive” elements of Go strategy by looking at previous games and learning from them. Schroepfer explains, “We’ve taken some of the basics of game-playing A.I. and attached a visual system to it, so that we’re using the patterns on the board—a visual rec[ognition] system—to tune the possible moves the system can make.”
Unsurprisingly, Facebook isn’t trying to build a better Go computer just for fun. They hope that the sophisticated program they’re developing for the board game will help them build more practical software in the future: Ultimately, they want to build A.I. that can make predictions based on existing evidence, and better mimic human intuition.
A "key problem in artificial intelligence is figuring out what’s going to happen next," Schroepfer told WIRED. “You do this all the time in order to make your day go well. … What we’ve got to do is teach computer systems to understand the world in a similar way.”