Why Google AlphaGo's victory over a human isn't just about Go

On Tuesday, Google’s artificial intelligence program AlphaGo completed a feat many thought would still be decades in the future – it beat human Go champion Lee Sedol in the complex, centuries-old game of strategy, winning four games to one.

The computer’s victory, which came at the end of a week-long showdown that was closely watched in Mr. Lee’s native South Korea and in China, marked a major milestone for artificial intelligence. Lee conceded at the end of a nearly five-hour-long final game, having won the previous match on Sunday.

“When it comes to psychological factors and strong concentration power, humans cannot be a match,” Lee said after the game concluded. But he added, “I don’t necessarily think AlphaGo is superior to me. I believe there is still more a human being can do to play against artificial intelligence.”

Beating a human at Go, which originates in ancient China, was long thought of as one of the most challenging tasks for artificial intelligence software, which has steadily made inroads into other games, include chess, Jeopardy!, and several Atari games, which Google’s software learned to play by analyzing the raw pixels onscreen.

The complex strategy game begins with one player using black stones while the other has white. The players take turns placing the stones on a 19x19 grid, each attempting to capture more territory on the board than the other.

Unlike chess or checkers, go pieces aren't moved around on the board. Players must capture an opponent’s stone by surrounding it with their own. That gives players an average of 200 possible moves for any particular position, compared with an average of 20 in chess, according to Demis Hassabis, the chief executive of Google’s DeepMind project.

That means Go has more possible positions than the the number of atoms in the universe, he wrote in a blog post in January after AlphaGo beat the European Go champion five games to zero.

The AlphaGo software draws on Google’s expertise in machine learning, a subset of AI that allows a computer to “learn” to complete particular tasks, and has been used for a range of applications, including identifying images, translating speech, or responding to emails.

The technique that Google used to crack the game combines a traditional artificial intelligence method where a computer figures out all the possible moves, then sorts through them to determine the best one. To do this, the computer must "see" all the way to the end of the game and form a search tree of possibilities to calculate if an individual move will help it win, The Christian Science Monitor’s Eva Botkin-Kowacki reported.

The Google researchers combined that with two deep neural networks that narrow down the game’s vast possibilities. One, known as the policy network, narrows the search to only include moves that are most likely to include a win, while the second “value network” evaluates if a move is stronger than others. This approach only sees far ahead enough to determine the immediate best move.

Mirroring the approach Google uses to “teach” its self-driving cars the rules of the road, AlphaGo was then put through a training method that combined studying moves made by human Go experts with playing millions of games against itself using the information from the two neutral networks.

That yielded what David Silver, the lead author of a paper the company’s researchers published in January in the journal Nature, has described as a “much more humanlike” approach to the complex game.

But the company says the purpose of AlphaGo isn’t simply to best some of the best players in the world, but to use the technology to expand the capabilities of artificial intelligence into other complex tasks.

While artificial intelligence technologies have also faced criticism – with physicist Stephen Hawking, entrepreneur Elon Musk, and others decrying the technology’s potential to be weaponized – researchers say computers’ ability to "learn" without the human problem of fatigue could have a number of applications.

In the short term, Google says it hopes to advance products such as smartphone assistants, while eventually using machine learning in applications such as healthcare and robotics.

“Because the methods we’ve used are general-purpose, our hope is that one day they could be extended to help us address some of society’s toughest and most pressing problems, from climate modeling to complex disease analysis,” Dr. Hassabis wrote in the blog post.

Researchers at Texas Tech for example, have been working on an effort to apply neural networks to climate modeling since 2012, while Facebook recently revealed it was using machine learning tools to develop highly detailed population maps that could be used to see how people connect to the Internet.

IBM’s competing Watson platform, which draws more extensively on cloud computing technology, is being used in a partnership with Memorial Sloan Kettering Cancer Center to help support physicians in diagnosing cancer cases in two hospitals in Thailand and India, tech site The Verge reports.

Watson, which uses what IBM calls “cognitive learning,” that relies more on predictive analysis than Google’s Alpha Go, is learning to recognize anomalies in medical images and flag them for a physician while learning to suggest possible treatments.

A more predictive approach could also be helpful for other areas, such as providing information about recidivism, or why people released from prison do or don’t commit additional crimes. That's a tool that researchers who have studied the technology say could be useful for judges making sentencing decisions, for example.

Google’s Go match may reveal some areas for the researchers to improve machine-learning technology.

Despite AlphaGo’s success in beating Lee, who was been making a living from playing Go since he was 12, the computer wasn’t completely infallible.

After he defeated the software during Sunday’s game, Lee noted that it didn’t handle surprise moves well. It also played less capably with a black stone, which needs to claim a larger amount of territory to win than an opponent with a white stone. In the final match, Lee chose the black stone, the Associated Press reports.

Other players pointed to the role of human fatigue in playing the game, a condition that didn’t affect Google’s computer, which could play hundreds of practices matches to sharpen its skills.

“It does not seem like a good thing for we professional Go players,” Chinese world champion Ke Jie told the AP on Tuesday, “but the match played a very good role in promoting Go.”