Google AI to play live Go match against world champion

[Bacchus]

From that basis I would suspect that the human merely played more solidly with extra time, but was equally hopeless.

That sounds plausible.

If the AI has a superior initial read of the position compared to the human ("instinctive" candidate move), I'm hard pressed to see how the human will catch up with extra time. Certainly not through calculation.

Well, here would be a good place for one of those tirades about the "beauty of Go". Calculation/reading is obviously highly important for Go, but it is definitely only a part of the game. There are all these questions about "right direction", and "what's important" that don't rely on reading much or in some cases at all. I obviously can't speak to how professional players answer these questions, but there is a lot of "feeling for what's right" by playing out the game in your head at a tactically basic level, and comparing roughly estimated outcomes on a holistic level. This presents a double challenge for computers -- they can't really play out the game "roughly", they have to actually look through the tree, and the trees for Go are large, in fact, they are so large, that whole-board trees are simply impracticable for calculation and will be for a decade or so; and computers aren't that good at holistic assessment. Basically, the extra time that's given the computer can use productively mostly to search through trees to optimize local plays — but that's normally a source of only a marginal advantage in Go. What the computer will not find with calculation (nor would a human) is a really big move that affects the structure of the board as a whole, especially as these moves can be away from the area where the immediate action is.

That's actually another story — understanding where is the immediate action? For a considerable part of the game, the action is split between completely separate theatres, which don't affect each other in any practically calculable way, but will eventually have some influence on each other. Finding the locally optimal move for the theatre is relatively easy for computers. What's more difficult is deciding which theatre gives a better return for a move right now, especially as taking that move on one theatre might well mean giving your opponent an extra move on another theatre. Even more difficult is accounting for how a locally optimal move will affect the theatres once they join up way, way down the line. Again something that privileges a big-picture view over move-by-move calculations.

[Gustaran]

Machines lead 1-0

http://money.cnn.com/2016/03/09/technolo...epmind-go/

Videostream of the first match:
https://www.youtube.com/channel/UCP7jMXS...KCAE0MHQ-A

Homepage:
https://deepmind.com/alpha-go.html

[picklepikkl]

The pro commentary seems to have settled on the analysis that AlphaGo was winning until about turn 80 (Lee played a very non-standard move early on, presumably to confuse a computer that might not have seen much of it, and then AlphaGo demonstrated that it was very able to cope), then Lee around moves 90-120, and then AG from 130 onwards (resignation after 186). AlphaGo would have won by about the komi -- the handicap given to white to make up for black getting the first move.

Very curious to see where this match goes, as an enthusiast of both AI and go!

[picklepikkl]

About twenty minutes ago, AlphaGo made a move in Game 2 that nobody at all understands. Lee Sedol left the room for 15 minutes to think about it, with his clock ticking down.

No report yet as to whether his ears turned red . What an interesting game! Guess we'll see soon whether it was an error or brilliance.

[Gustaran]

AlphaGO won the second game as well.

If anyone is interested in learning the Go rules, this site features nice interactive instructions:

http://playgo.to/iwtg/en/

[Bacchus]

An absolute rout, I would say, the question is whether Sedol manages to win even a single game here. The AI is really brilliant, I sort of hope for a human all-stars match against it, as unlikely as that is.

[Gustaran]

And it's 3-0 - machine wins!

If anyone is interested, before the game today AlphaGo's lead project manager answered some interesting questions on how the software works:

https://youtu.be/qUAmTYHEyM8?t=15m24s

[NobleHelium]

I quite like the American's commentary. Much better than your normal sports commentary.

[Bacchus]

He's done it! Absolutely amazing stuff, an insanely complicated tactical play arising out of a strategic preparation that took a dozen or so turns at different parts of the board. Humans are good for something after all

[NobleHelium]

He also described the so-called "god move" as "the only move [he] could make" in the press conference. Not sure what to make of that.