Google's Go-playing computer program AlphaGo, which incorporates artificial intelligence, convincingly beat South Korean Go grandmaster Lee Se-dol 4-1 in their recent five-game showdown.
Chess and shogi had already surrendered to powerful computer software and the ancient game of Go was seen as the last bastion in the battle against machines. But it, too, has now capitulated. Go is a perfect information game, meaning it is not swayed by chance and in theory there is a certain path to victory, so humans were bound to be overpowered sooner or later. However, it was widely believed that that day was still at least 10 years away. The unexpected victory by Google's software was, to borrow an example from Japanese history, like the sudden arrival of the Black Ships. We can even say that the program's win is the biggest event in the 4,000-year history of the game of Go.
There were, of course, signs that this would happen. On Jan. 28, a paper was released, stating that AlphaGo had beaten a professional Go player in a match without handicaps. The program defeated a Chinese player with a second-dan rank. In October last year it was said to have won all five games of a contest, but the match was against a player I had never heard of, and to tell the truth, it felt unreal to me.
But upon examination of the moves of the games, I was surprised. The program was clearly a step ahead of other Go programs that had surfaced before then. Previous software programs had a characteristic tendency to suddenly make cryptic moves, and it was easy to pick out that the game was being played by a piece of software. But AlphaGo didn't share that tendency. Professional players who examined game data were united in the view that it was impossible to tell whether the moves were made by a human or a software program. In other words, one could conclude that AlphaGo had already by this stage achieved a level of skill similar to that of a human.
Then in March, AlphaGo was pitted against South Korea's Lee Se-dol, who holds a ninth-dan ranking. They played five games against each other.
The software indeed possessed a degree of competency, but I thought, "There's no way it can suddenly become strong in around half a year," and predicted that Lee would achieve a sweeping victory, not losing any of his games. To tell the truth, I had pride as a Go player myself, having played for about 50 years and holding a seventh-dan rank. But when the curtains were raised and the two sides got going, it turned out to be a "one-man show" -- by AlphaGo.
The program achieved an easy win in the first game, seizing the initiative from the opening, and playing powerful moves in the middle game. The second game started out with a unique structure and the software program maintained control of the whole game. In the third game, too, the computer program secured the upper hand with a brilliant series of moves in the opening, and while it faltered slightly in the endgame, it managed to escape with the win to make the score 3-0. In the fourth game, AlphaGo went off track from the middle game and lost like other Go software programs. But in the final game, it played faultlessly, eventually forcing Lee to resign.
The biggest reason for the software program's breakthrough was its ability to secure a "wide perspective" of the game. Software programs to date had a strong tendency to rely on the power of their calculating speed in reading moves. But AlphaGo amassed knowledge through the process of "deep learning," at the forefront of computer technology. More than 100,000 games were inputted, and the software memorized as many as 30 million positions. It also played many matches itself and is said to have learned what patterns and stone placements were advantageous, and how to convert these into wins. This is what humans would call a wide perspective or an intuitive feeling for the game. It's like a young child watching an adult play many times, and naturally coming to understand the rules and become stronger.
On March 19 and 20, the University of Electro-Communications in the Tokyo city of Chofu held a tournament in which Go programs faced off against each other. A record high of 31 teams took part, and the program "Zen," developed by Japanese programmers, claimed victory for the first time in two years and the third time overall. The runner-up was a software program created by a Facebook researcher. On March 23, the top two software programs went up against Go player Koichi Kobayashi, who holds an Honorary Kisei title, being allowed a three-stone handicap. The runner-up program was defeated, but Zen maintained the advantage throughout the game and overcame the Go master. This showed that the level of Japan-developed software programs was certainly improving.
But when software programs become stronger and overtake humans, will this take some of the shine off the appeal of Go?
Included among AlphaGo's moves were ones that had been considered "bad" according to human common sense to date. But if software programs produce a new form of common sense through their big picture of the game and humans learn from it, then the world of Go will probably expand.
It is hoped that Google will publicly reveal how AlphaGo evaluated the various positions as it proceeded through its games.
The rules of Go, which uses black and white stones, are extremely simple. But the game has a depth and beauty that stems from virtually unlimited possibilities. Because of that, the game has fans all over the world and the playing population is said to number 40 million. If people can take something from the power of software and use this to boost the game's appeal, then Go will probably continue to hold up throughout the ages and beyond generations. (By Moriei Kanazawa, Cultural News Department)