Google's AlphaGo AI To Fight World's Go Champion Tonight

YouTube will stream live the five games between Google’s AlphaGo AI and the world champion at Go, Lee Sedol. It has been said for decades that true AI will arrive when it can finally beat humans at Go, and Google’s AI may be able to achieve that soon. The first game will begin today at 11pm ET on YouTube. The next four will follow over the next few days.

Go is a 2,500 year old game, played mostly in Asia, and it requires a high level of intuition as well as discipline from the player. The intuition requirement has made creating a Go AI so challenging in the past because it’s one of the human capabilities that machines haven’t come close to replicating well enough yet.

Intuition is sort of a “shortcut” our brain uses to calculate many possibilities in a split second and then give us a “good enough answer” to a certain problem that needs to be solved quickly. Until recently, AIs hadn’t been built like that. Instead, they had to calculate all the possible moves for a certain number of steps into the game. Then the AI would pick the best move out of those.

This is how chess AI has worked, for instance, as when IBM’s Deep Blue beat Garry Kasparov. However, unlike chess, which can have 20-30 possible moves at any given time, Go players have about ten times as many options. This quickly makes Go a much more complex game for a machine, because all of those possibilities need to be multiplied at each step. This means that a Go AI can’t beat a human by trying to “bruteforce” all potential future moves and then picking the best one.

With AlphaGo, Google has “taught” the AI 30 million games of Go, from which the AI “learned” how to play well, much like a human would. This gave AlphaGo a 57 percent chance to predict a human move. However, this meant it could still be beaten by humans. To improve it further, Google matched AlphaGo against itself, each version trying to beat the other, and each learning from the other’s mistakes and successes, until it evolved much further.

Earlier this year, AlphaGo managed to beat the European champion at Go, in five matches out of five. However, the world Go champion is supposed to be a much better player because Go is very popular in China, so he's been training with great players since childhood.

Chances are that Google’s AlphaGo will still end up beating him, because it’s likely that the AI has already improved dramatically since he beat the European champion 5-0. That’s how computers work, and even if Lee Sedol manages to beat it now at least 3 times out of 5, it's probably only a matter of time before the AI catches up.

Lucian Armasu is a Contributing Writer for Tom's Hardware. You can follow him at @lucian_armasu. 

Follow us on FacebookGoogle+, RSS, Twitter and YouTube.

Lucian Armasu
Lucian Armasu is a Contributing Writer for Tom's Hardware US. He covers software news and the issues surrounding privacy and security.
  • hellwig
    You referred to the AI as "HE" in the last paragraph. Nice TRY AlphaGo! You almost passed the Turing test.

    Seems to me you could "easily" solve all these types of games with a relational database. If Google can index and archive millions of web sites, and poll that data and respond to web queries in fractions of a second, why couldn't they develop a database of every possible state of the board, and simply traverse that like they do the database of websites? I know that wouldn't be AI in any sense, just a machine finding the next most likely state, but it seems to me that Google has the resources at hand.
  • ThatsANoGo
    They couldn't do that because a Go board has 19x19 intersections, each of which can be either blank, or have a white or black stone on it. This gives a grand total of 3^361 possible configurations before we consider the actual configurations that are legal, which is "only" about 2*10^170.

    It's impossible to convey just how stupendously enormous that number is. It's far, far bigger than the number of elementary particles in the universe. Actually, if every elementary particle in the universe held its own universe inside it somehow, then the total number of nested particles would still be great way off from this number.

    So no, Google definitely does not have the resources at hand. God himself doesn't have the resources at hand. Of all the ways you could have a computer "solve" the game of Go, this one is definitely, capital-I impossible.
  • voljin
    To give you a taste of just how large that number is have a look at which is a representation of ~8*10^67, a MUCH MUCH MUCH smaller number
  • bwohl
    Say hello to your new robot overlord.....
  • From what I "understand", Google's "solution" to the problem is basically self-evolved software. While neat, it doesn't give greater fundamental insight into intelligence or estimation or strategy. It's basically a black box that works, but we don't really know how.
  • ErikVinoya
    Machine learning algorithms have existed for quite a while now. Recursive neural networks perhaps?
  • 17627681 said:
    Machine learning algorithms have existed for quite a while now. Recursive neural networks perhaps?

    Bayesian even, maybe. Still, the match is awesome so far.
  • gangrel
    To give some idea of how massive...

    1 light year is basically 10^16 meters. 1 angstrom, the length unit for light wavelength, is 0.1 nanometers, or 10^-10 meters. Ergo, 1 light year is 10^26 angstroms. Space is believed to be 10^10 years old, and therefore probably 10^10 light years in diameter, or 10^36 angstroms.

    So, the volume of known space, measured in *angstrom units* about 10^110.

    Still short 60 decimal places.
  • norby_7f
    To give you some idea of how massive...

    Chuck Norris says it's pretty hard to count to 10^170; and he counted to infinity, twice...
  • poochiepiano
    This might be hard to answer, but I would imagine that each successive move would temp down the probabilities significantly. I don't know if this approach, if possible, would be better or worse than whatever Google is doing, but would it even be feasible to factor this in to the late-game moves where options are much more limited?