Google’s AI Program AlphaGo Zero Has Mastered Chess in Just 4 Hours

Fagjun | Published 2017-12-08 22:33

AlphaGo Zero, the board game-playing artificial intelligence program by DeepMind, has conquered chess in just a matter of a few hours.


AlphaGo Zero is dominating chess.

AlphaGo Zero is dominating chess.

 

 

This isn’t the first time that AlphaGo Zero has managed to master a board game. The program’s earlier incarnation, AlphaGo, was able to beat the ancient Chinese board game Go’s world champion, Ke Jie. AlphaGo learned to play the game by watching hours upon hours of gameplay, but AlphaGo Zero was able to teach itself how to play Go as well as other games by repeatedly playing against itself.

 

It seems that AlphaGo Zero is determined to keep impressing. The program taught itself the rules, techniques, and strategies of chess in less than four hours. Afterward, AlphaGo Zero was able to defeat the chess-playing program Stockfish 8 in a performance that researchers have called “superhuman”. While there’s already a study on the program, however, it has yet to undergo peer review.



The Superhuman Artificial Intelligence


AlphaGo playing against Lee Sedol [Photo by Getty Images]

AlphaGo playing against Lee Sedol [Photo by Getty Images]

 

 

Stockfish 8 was first released in 2008, and volunteers spent the next two years building upon the program. It also previously won the Top Chess Engine Championship in 2016, though it was beaten by Komodo, another program, twice this year.

 

AlphaGo Zero played a series of 100 games against Stockfish 8, culminating in no wins for Stockfish and no losses for AlphaGo Zero. AlphaGo Zero won 25 games playing as white and three games playing as black, while the rest of the games ended in a draw. Both AlphaGo Zero and Stockfish were given a minute each to figure out their each move.

 

What’s even more impressive is that the AI managed to learn by itself. “It’s a remarkable achievement, even if we should have expected it after AlphaGo,” Garry Kasparov, a former world chess champion, said in an interview. “We have always assumed that chess required too much empirical knowledge for a machine to play so well from scratch, with no human knowledge added at all.”

 

Kasparov isn’t the only chess master that seems to be impressed with the results. "I always wondered how it would be if a superior species landed on earth and showed us how they played chess," said Peter Heine Nielsen. "Now I know."



Learning Without Humans


Deep Blue versus Garry Kasparov, May 12, 1997. [Photo by Stan Honda/AFP/Getty Images]

Deep Blue versus Garry Kasparov, May 12, 1997. [Photo by Stan Honda/AFP/Getty Images]

 

 

Google has declined to comment on AlphaGo Zero’s achievements in chess before its publication in a journal. However, there is a paper on Arxiv about an algorithm called AlphaZero that has learned how to play chess in four hours and has beaten Stockfish. The program, according to the paper, was only given the basic rules of chess before learning the game by playing against itself.

 

The paper adds that chess wasn’t the only game that AlphaZero was able to master. It was also able to beat the competition, another AI program named Elmo, at a Japanese board game called Shogi. AlphaZero won 90 games, lost eight, and ended two in a draw. AlphaZero was also apparently able to beat its previous incarnation at Go, with 60 wins and 40 losses.

 

However, human chess masters have been beaten by machines before. IBM”s Deep Blue, for example, defeated Kasparov in 1997. The difference between AlphaZero and other programs, DeepMind says, is that AlphaZero learned the game by itself.

Hey! Where are you going?? Subscribe!

Get weekly science updates in your inbox!