Wednesday, March 16, 2016

Will the Go Algorithm Use the Million Dollar Prize to Enter the World Series of Poker?

Following up on this morning's "So, What Will Google’s Winning Go Algorithm Do Now That It's Won the Million Bucks?".

From the Los Angeles Times:

A computer is now the Master of Go -- but let's see it win at poker
The worlds of Go and artificial intelligence were both unsettled by the victory this week of an advanced artificial intelligence computer over one of the world's leading masters of the intricate Japanese board game, four games to one. It's an achievement that experts in both fields didn't expect to happen for as long as 10 years.

The triumph of AlphaGo, the product of a Google lab named DeepMind, over the fourth-ranked Go champion, Lee Sedol of South Korea, is widely viewed as a landmark in artificial intelligence much greater the victory of IBM's Deep Blue over chess Grandmaster Garry Kasparov in 1997. That result, Kasparov wrote in 2010, "was met with astonishment and grief by those who took it as a symbol of mankind’s submission before the almighty computer." Go is a far more complex challenge than chess, so it's unsurprising that AlphaGo's victory is seen as bringing the era of ultimate submission that much closer.

But is it? Some AI experts are cautious. “People’s minds race forward and say, if it can beat a world champion, it can do anything,” Oren Etzioni, head of the the nonprofit Allen Institute for Artificial Intelligence in Seattle, told the journal Nature this week. The computational technique employed by AlphaGo isn't broadly applicable, he said. “We are a long, long way from general artificial intelligence.”

And there's another aspect to consider. Chess and Go are both "deterministic perfect information" games, Alan Levinovitz of Wired observed in 2014 -- games in which "no information is hidden from either player, and there are no built-in elements of chance, such as dice." How will a computer do in a game in which key information is hidden and the best players win by using the unique human skill of lying? In games such as poker, Science magazine points out, in which victory depends not on pursuing the optimal strategy, but deviating from it?

First, a few details of the game at the center of this week's match. Invented more than two millennia ago in China, Go is played on a grid of 19-by-19 lines on which two players place black or white stones, attempting to build chains that surround territory without being enveloped by their opponent. Because of the size of the board and the lack of specific rules for each piece, the number of potential moves in Go is exponentially larger than in chess.

The number of possible arrangements of stones is on the order of 10 to the 100th power; far more than the options that Deep Blue had to consider in defeating Kasparov. The route to victory at any point can be obscure; experts talk as though their play emerges from intuition, or even the subconscious, as much as experience. Indeed, during Sedol's lone victorious game, DeepMind chief Demis Hassabis tweeted that AlphaGo hadn't played a wrong move, but had become deluded into believing it was winning. Go occupies a special position in oriental life; in his classic work, "The Master of Go," the Nobel Prize-winning novelist Yasunari Kawabata spun a trenchant tale of youth versus age, the past versus the present, and the power of culture out of the game.

The DeepMind designers' solution to the intricacy of Go was to use an architecture known as neural networks. These mimic the structure of the human brain by creating connections that become stronger with experience -- in other words, to learn. AlphaGo could test millions of options and assess their outcomes rapidly, but its design allowed it to develop shortcuts to discard all but the most promising choices. Observers found the result strikingly human-like; the machine's move  No. 37 in game two was so unexpected that some commentators thought it was a mistake. Sedol left the room in shock, and later confessed, "Today I am speechless." See the move and the commentators' reactions below, at 1:18:25 of the recording:...MORE
The artificial intelligence folks still at Carnegie-Mellon after Uber bought the autonomous vehicle department have figured their road to riches is not Travis Kalanick but online Limit Hold-em.
(kidding, research purposes only. they say)