How difficult is Go

AI milestone: The game of Go is cracked

In the end, things went much faster than expected. "In ten years", experts have answered so far when one asked when a computer would be able to defeat top human players in the Go. The board game, which originated in East Asia, is notorious for its complexity - which is why it has long been an important milestone in the further development of artificial intelligence (AI). And now it has happened: A team from the Google research laboratory "DeepMind" in London reports that their system has beaten the reigning European champion Fan Hui with 5 to 0 games - one might add: devastatingly.

Unlike chess, in which even grandmasters have no chance against computers today, Go has so far resisted artificial intelligence and was therefore one of the last bastions of human superiority in the game. What the AI ​​researchers lacked above all was a sophisticated solution - and Demis Hassabis' team has now found it in the technology of so-called deep learning. A process that is currently cracking one AI problem after another.

This article is contained in Spectrum - The Week, 05/2016

"AlphaGo", as Hassabis and colleagues called their software, played like a human, says Toby Manning of the British Go Association, who oversaw the game against Hui: "You couldn't say who was who during the game." AI researcher Jonathan expects that the success will cause enthusiasm in research circles: "Nobody expected such a leap forward," he explains in the "New Scientist".

In Go, players place pieces of their color (black or white) one after the other on a 19 by 19 board. It is important to surround opposing stones and thereby conquer them for yourself. Whoever controls more than half of the board wins. In games of this kind, in which chance does not play a role, it makes sense to let the computer calculate all possible future moves in advance - until the end of the game. The result is similar to a tree that branches out for each move by the number of all valid subsequent moves.

Why chess is so difficult - and Go even harder

If such a tree can be set up, the computer simply has to select the moves that lead to a win in the game. Unfortunately, it is not possible with chess, and certainly not with Go, to follow all the branches of this search tree. Their number quickly exceeds the limits of any manageability. But the whole thing can also be solved much more easily if you only look at those paths that are particularly rewarding and, secondly, only follow them step by step into the future until you know for sure enough whether the game is going in a favorable direction developed. Unnecessary and hopeless paths are simply sorted out. The only difficulty is: How do you know that you should stop? And how do you even find the worthwhile moves?

Thanks to differently weighted pieces and a smaller playing field, the game of chess provides more clues, which ultimately allowed an IBM team to defeat the chess grandmaster Garry Kasparov with their program "Deep Blue" in 1997. Go, on the other hand, makes it particularly difficult for an AI. There are no formal differences between the stones, and whether it was useful to place one of them in a certain space often only becomes apparent much later in the game. The "value" of a given position is therefore difficult to determine.