#HTE

A Computer Has Beaten a Human Champion at Go. What Next?

Sometimes it’s fun to look back at old predictions of technological progress and compare them to present-day reality.

Other times it’s a little terrifying.

On May 12, 2014—less than two years ago—Wired magazine ran an article headlined “The Mystery of Go, the Ancient Game That Computers Still Can’t Win.” It described the ongoing—and, to that point, fruitless—quest by artificial intelligence researchers to build a computer program that could defeat a top human player at Go, a board game that’s orders of magnitude more complex than chess.

The problem: The number of possible moves and configurations in Go is too great for any existing computer to fully analyze. Humans can’t fully analyze them either, of course—which is why the top players rely heavily on intuition, which has never been machines’ forte. The top program at the time, called Crazy Stone, had “no chance” of beating a human champion, its creators admitted. Asked to speculate on how long it might be before anyone accomplished the feat, they offered a guess of 10 years, a number that Wired warned “may prove too optimistic.”

It did not prove too optimistic.

On Wednesday, Google announced that AlphaGo, a program built by its DeepMind artificial intelligence lab, had defeated the European Go champion, without a handicap, in a five-game match last October. The score: 5-0.

The next test will come in March, when AlphaGo challenges the world champion, Lee Sedol, in a five-game match in Seoul. Lee is a legendary master of the game, and by consensus the best player of modern times. It will be the Go equivalent of the famous chess match between IBM’s Deep Blue and Garry Kasparov in 1996.

Google broke the news in an article published this week in the journal Nature, titled, “Mastering the game of Go with deep neural networks and tree search.” The DeepMind team’s approach combined a well-established algorithmic approach known Monte Carlo tree search, which has helped computers defeat humans at many less-complex games, with a cutting-edge approach known as deep neural networks. AlphaGo actually uses two separate neural networks: a “policy network” that limits the scope of its analysis to a handful of attractive options for each move, and a “value network” that peers about 20 moves into the future to see which of those options appears the most promising. The team explains how this works in the Nature video below:

The implications of Google’s achievement stretch beyond the realm of games. Neural networks can be used in all sorts of settings that demand the human-like capacity to evaluate various strategies under conditions of uncertainty. Virtual assistants and medical diagnostics come to mind, and Google mentioned climate modeling in a blog post. At this point, to identify all the possible applications of such a fundamentally potent technology would require a feat of imagination in its own right. It’s no wonder that Facebook’s top AI researchers are hard at work on the very same problem. (They’re a little behind.)

And yet it would be a mistake to conclude, even half-jokingly—as many were quick to do in the wake of Google’s announcement—that machines are now intellectually superior to humans, or anywhere close, really. As complex a game as Go is, it isn’t really “a microcosm of the real world,” as DeepMind founder Denis Hassabis claimed to the New York Times. It is a constrained environment, albeit a vast one, in which both contestants have access to perfect information, albeit too much information to fully process, and share the same perfectly defined goal. Real life is simply not a game in this sense, and no algorithm yet devised could begin to approach the mental flexibility required to navigate it in the way that humans do.

I’ll hold off on predicting that it will never happen, though. At this rate, just about anything is possible.

In the meantime, there remain other games and tasks for computers to conquer. Next on the agenda after Go might be No Limit Texas Hold ’Em poker. A software program in 2014 solved two-person limit hold ’em, but humans still reign at no-limit—for now.

Nor has Go been “solved,” exactly. AlphaGo is really good at the game, but it isn’t perfect, nor is it designed to be. And for what it’s worth, South Korea’s Lee is predicting victory in his March match against AlphaGo. “I heard Google DeepMind’s AI is surprisingly strong and getting stronger, but I am confident that I can win, at least this time,” he said in a statement.

He may be right. People sometimes forget that Kasparov won that 1996 match, 4-2. That’s probably because Deep Blue went on to beat him in a rematch the very next year.

Previously in Slate:


http://www.slate.com/blogs/future_tense/2016/01/28/a_computer_google_s_alphago_has_beaten_a_human_go_champion_what_s_next.html