AlphaGo AI stuns go community

Google’s artificial intelligence program AlphaGo stunned players and fans of the ancient Chinese board game of go last year by defeating South Korean grandmaster Lee Sedol 4-1. Last month, an upgraded version of the program achieved a more astonishing feat by trouncing Ke Jie from China, the world’s top player, 3-0 in a five-game contest. In the world of go, AI appears to have surpassed humans, ushering in an age in which human players will need to learn from AI. What happened in the game of go also poses a larger question of how humans can coexist with AI in other fields.

In a go match, two players alternately lay black and white stones on 361 points of intersection on a board with a 19-by-19 grid of lines, trying to seal off a larger territory than the opponent. It is said that the number of possible moves amounts to 10 to the power of 360. This huge variety of options compels even top-class players to differ on the question of which moves are the best. Such freedom to maneuver caused experts to believe it would take a while before AI would catch up with humans in the world of go. Against this background, AlphaGo’s sweeping victory over the world’s No. 1 player is a significant event that not only symbolizes the rapid development of computer science but is also encouraging for the application of AI in various fields.

In part of the contest with Lee in Seoul in March 2016, AlphaGo made irrational moves, cornering itself into a disadvantageous position. But in the case of its contest with Ke in the eastern Chinese city of Wuzhen in late May, it made convincing moves throughout, subjecting the human to a “horrible experience.” He called AlphaGo “a go player like a god.”

AlphaGo was built by DeepMind, a Google subsidiary. It takes advantage of technology known as deep learning, which utilizes neural networks similar to those of human brains to learn from a vast amount of data and enhance judging power. This is analogous to a baby learning a language by being exposed to a huge volume of utterances over a period of time. The program not only learns effective patterns of moves for go by studying enormous volumes of documented previous games but also hones its skills by playing millions of games against itself. In this manner, it has accomplished a remarkable evolution over the past year. Unlike humans, it is free of fatigue and emotional fluctuations. Because it grows stronger by playing games against itself, there is no knowing how good it will become in the future.

Feeling intimidated by AI programs should not be the only reaction of human go players. They can receive inspiration from AlphaGo since it shows a superior grasp of the whole situation of a contest, instead of being obsessed with localized moves, and it often lays stones in and around the center of the board. Human players usually first try to seal off territory around the corners. Its playing records also prove that even some moves traditionally considered as bad have advantages. By learning from AlphaGo, go players can acquire new skills and make their contests more interesting.

AlphaGo does have a weak point. It cannot explain its thinking behind the particular moves that it makes. When watching ordinary go contests, fans can enjoy listening to analyses by professional players. Also, ordinary go contests are interesting since psychology plays such an important part of the game, especially at critical points. This shows there are some elements of go that AI cannot take over.

DeepMind is thinking about how it can apply the know-how it has accumulated through the AlphaGo program to other areas, such as developing drugs and diagnosing patients through data analysis. But the fact that the program made irrational moves during its match with South Korea’s Lee shows that the technology is not error-free — a problem that must be resolved before AI can be applied to such fields as medical services and self-driving vehicles. Many problems may have to be overcome to make AI safe enough for application in areas where human lives are at stake.

A report issued by Nomura Research Institute says that in 10 to 20 years, AI may be capable of taking over jobs now being done by 49 percent of Japan’s workforce. At the same time, it says AI cannot intrude into fields where cooperation or harmony between people is needed or where people create abstract concepts like art, historical studies, philosophy and theology. It will be all the more important for both the public and private sectors to make serious efforts to cultivate people’s ability to think and create while finding out what proper roles AI should play in society.