Will Go AI Eclipse Top Human Gamer This Week?

Having beaten Europe's top Go professional in October, Deepmind's AlphaGo computer program is now set to face the world's top Go professional this week in a five game match for $1 million stakes. Will the solar eclipse hours before the match starts affect the outcome?

Kasparov vs. Deep Blue

Jim Gardner / Via Flickr: 22453761@N00

Just over 20 years ago, IBM's Deep Blue became the first chess computer program to defeat a reigning chess world champion by winning its first game against Gary Kasparov on February 10th, 1996. Kasparov proceeded to win the match, but then lost a match to Deep Blue in 1997. I predict AlphaGo's best shot at beating Lee Se-dol this week will come in the first game as well. AlphaGo has much more information about Se-dol's games and play style than vice versa.

Chess vs. Go

Via siansullivan.net

The ancient oriental game of Go has much simpler rules than chess, but due to the larger size of the Go board (19x19 vs. 8x8) and number of possible moves in a Go game, there are astronomically more possible Go positions than chess positions. The exact number of legal Go positions - a 171 digit figure - was actually calculated down to the last digit (5) earlier this year, after several months of computation. That number leaves the estimated number of valid chess positions, which is less than 48 digits, far behind. Due in part to this large disparity, until recently, Go playing computer programs needed a handicap of several stones in order to prevail against higher level professional human players. Given recent advances in computer Go play and AlphaGo being able to solve a mathematical problem about Go (estimating the eventual winner given any Go position) which was believed to be too difficult to solve in 2002, a computer Go program will likely overcome the top human Go player within under 5 years. Humans pwned? Not quite yet...

Go vs. MTG

Magic: the Gathering has been shown to be a more complex game than either chess or Go. Granted, as MTG is currently played it is not a perfect information game which minimizes chance, but... MTG games could be played between humans and computer programs with all cards visible at all times to both players and with deck shuffling cards eliminated from the decks used in order to allow for both perfect information and minimized randomness. If either player won against the other player playing from both sides of an identically shuffled pair of decks, the winner would clearly be the stronger of the two players. Games with less complex rules than MTG which would be more complex than Go and yet allow for perfect information and minimized chance could also be developed.

AI vs. 8th Grade science tests

However, even if within the next decade or so AI can pwn humans at any tabletop game, it still would flunk 8th grade science classes at its current level of development, so all of us HS grads can take some small comfort from that fact for perhaps the next decade or two.

AlphaGo vs. Lee Sedol

youtube.com

The Lee Sedol vs. AlphaGo match will be streamed live on youtube and there will be professional level commentary provided in English by Michael Redmond. The match, however, is not at a convenient time for the nine to fivers on the east coast of the U.S., unfortunately.