Evonne/Flickr

Machine learning can finally beat humans at this ancient Chinese game

Eventually, the machines always win.

Mar 1, 2020, 2:38 pm*

Tech

 

Taylor Hatmaker

Once again, humans are on the brink of falling to our future robotic overlords. When it comes to man vs. machine, mastery of the ancient Chinese game of Go remained squarely in human territory—robotic challengers might put up an okay fight, but they’d yet to best the strange ingenuity of the human mind… until now.

AI systems are already good at stuff like chess and Spike Jonze movies, but perfecting the art of Go, an ancient Chinese strategic board game dating back 2,500 years, presented some unique challenges for non-human players. If you’re not familiar with the game, it’s played like this:

One major computational challenge is that the number of potential moves at any stage of the game often numbers in the hundreds—far more options for an algorithm to consider than a game like chess might present. Computers also struggle to figure out who has the upper hand at any given point in the game. Given the game’s deep strategy, the number of black or white “stones” remaining on the board don’t necessarily indicate which player is winning. 

For better or worse, the University of Edinburgh’s Christopher Clark and Amos Storkey are working around those challenges:

“Mastering the game of Go has remained a long standing challenge to the field of AI. Modern computer Go systems rely on processing millions of possible future positions to play well, but intuitively a stronger and more ‘humanlike’ way to play the game would be to rely on pattern recognition abilities rather then brute force computation.”

According to the write up in MIT’s Technology Review, Clark and Storkey’s new Go-playing AI system draws on a database of 16.5 million states of the board and potential moves, all culled from actual Go matches between expert players. The computers were trained to crunch these numbers to determine what move either player might make next at any given time. The result was that the AI correctly guessed the expert player’s next move 44 percent of the time— a major leap in algorithm-powered Go skill. 

Naturally, the most interesting and unsettling part is that the AI upped its Go game by learning to think more like a human. Our days of sentient superiority are clearly numbered.

H/T MIT Technology ReviewPhoto via Evonne/Flickr (CC BY-2.0)

Share this article
*First Published: Dec 16, 2014, 2:59 pm