Computer Go

[Christopher Kreuzer]

BBC article here:

http://www.bbc.co.uk/news/technology-35746909

AlphaGo playing Lee Se-Dol.

Does anyone know if the results are being announced anywhere 'live' (for all I know there is a live Go site somewhere, but I doubt it somehow - I suspect live transmission on YouTube will be what is happening).

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

"The matches will be held at the Four Seasons Hotel, Seoul, South Korea, starting at 1pm local time
(4am GMT; day before 11pm ET, 8pm PT) on March 9th, 10th, 12th, 13th and 15th."

I think that means the first game is at 4am on Wednesday morning (9th March).

"The matches will be livestreamed on DeepMind’s YouTube channel"

Live stream for game 1 is here:

https://www.youtube.com/watch?v=vFr3K2DORc8

Time control is interesting (compared to chess time control at similar levels):

"Each player will receive two hours per match with three lots of 60-second byoyomi (countdown periods after they have finished their allotted time). Each match is expected to take around 4-5 hours."

I have no idea what three lots of 60-second byoyomi means? I know that in Go matches between humans, the game seems to finish far earlier than I would, because I don't understand that the rest of the moves will just be filling in the gaps and there is no more 'territory' to be won (i.e. the existing structures on the board are safe). Wonder how a computer handles that, does it just carry on right to the end, maybe catching humans out sometimes in 'simple' positions?

(Apparently AlphaGo still sometimes makes elementary mistakes. And is expected to lose the match, but may win a game, which would itself be a landmark - the earlier win was against a much weaker player.)

[MartinCarpenter]

It was the European champion. It really isn't a European game

Dunno what they'll do at the end of the games. Think its mostly quite easy to identify absolutely safe structures?

Anyway, lets hope we can keep on top in this for say 5 years.....

[Paolo Casaschi]

Very sad to see that the go match between a computer and the best human gets a lot more attention on the media than the upcoming chess candidates tournament and any recent chess event, world championship match included.

[MartinCarpenter]

What, why? Its really quite a major event on various grounds.

Go is a major board game in world terms - very big in China/Japan etc - and the AI advances they've had to produce to play it well are actually much more inherently interesting than the steps required to win at chess.
(Which basically boiled down to the slightly smarter application of increasing amounts of brute force!).

Actually, there are quite major medium term social implications here. Its more of a symptom than a cause of course, but if they can play Go this well, then there's a lot of things that we do now that can be automated very effectively.

[Paolo Casaschi]

Which basically boiled down to the slightly smarter application of increasing amounts of brute force!

That's where the media exaggeration comes into place: I read few reports mentioning chess applications being based on brute force and this new go application being based on something completely different. I do not know about go but as far as I know, chess engines shy away from brute force as much as possible; in fact every software breakthrough seems to come from new techniques distancing even more from brute force. Also, many of the keywords mentioned for the go application have been already used in chess, for example Montecarlo simulations to assess the best move and engine tuning using games against the engine itself.
I do not want to minimize the recent achievements with go, but I'm not sure whether those achievements are based on a revolutionary approach or just the standard google approach to throw a lot of money and a LOT of computing power onto a problem.

[David Shepherd]

I have no idea what three lots of 60-second byoyomi means?

It means you have 60 seconds per move which you can exceed twice but once you exceed it the third time you lose. If you exceed the 60 seconds on an individual move you start using the next 60 seconds so you could lose all the byoyomi periods and the game if you spent over 180 seconds on a move.

[Brian Towers]

I have no idea what three lots of 60-second byoyomi means? I know that in Go matches between humans, the game seems to finish far earlier than I would, because I don't understand that the rest of the moves will just be filling in the gaps and there is no more 'territory' to be won (i.e. the existing structures on the board are safe). Wonder how a computer handles that, does it just carry on right to the end, maybe catching humans out sometimes in 'simple' positions?

That final stage is far and away the easiest. It is getting there that is the tricky part.
When you are playing a stronger player he will give you a very strong hint that you have reached the end of the game by passing. You are, of course, free to continue playing stones but the risk you take is that you are not placing them in dead space but actually reducing your own territory.

[Christopher Kreuzer]

I have no idea what three lots of 60-second byoyomi means? I know that in Go matches between humans, the game seems to finish far earlier than I would, because I don't understand that the rest of the moves will just be filling in the gaps and there is no more 'territory' to be won (i.e. the existing structures on the board are safe). Wonder how a computer handles that, does it just carry on right to the end, maybe catching humans out sometimes in 'simple' positions?

That final stage is far and away the easiest. It is getting there that is the tricky part.
When you are playing a stronger player he will give you a very strong hint that you have reached the end of the game by passing. You are, of course, free to continue playing stones but the risk you take is that you are not placing them in dead space but actually reducing your own territory.

Thanks, that is what I largely remember. I was probably insulting the stronger players I was playing by playing on in those situations. I think I would also pass before that point was reached, which would prompt polite constructive criticism after the game. I was also never able to actually keep track of how much territory I had and whether I was ahead or behind overall. Bit of a handicap that!

Is there an etiquette in Go regarding passing - such that you shouldn't try and bluff around things such as passing? And again, I wonder how that is handled with computers. Is there a 'handler' for AlphaGo that will pass or resign on behalf of the computer?

Here's a question: if the board in Go was expanded beyond the current 19x19, to say 23x23, would that favour humans over computers again, or not? At what point does expanding the board become silly?

[Paolo Casaschi]

Each player will receive two hours per match with three lots of 60-second byoyomi (countdown periods after they have finished their allotted time). Each match is expected to take around 4-5 hours.

I do not want to sound more pedantic than usual, but if each player has 2 hours per game (plus a few 60 second bonuses), how can possibly a game last 5 hours? How civilized if they allowed for a tea break in the middle of the game

[Christopher Kreuzer]

Each player will receive two hours per match with three lots of 60-second byoyomi (countdown periods after they have finished their allotted time). Each match is expected to take around 4-5 hours.

I do not want to sound more pedantic than usual, but if each player has 2 hours per game (plus a few 60 second bonuses), how can possibly a game last 5 hours? How civilized if they allowed for a tea break in the middle of the game

The 60-second bonuses are regenerated if you don't use them up: "It means you have 60 seconds per move which you can exceed twice but once you exceed it the third time you lose."

http://www.britgo.org/bgj/10643.html

Nice description of a human timekeeper counting out the time there!

This bit may also be relevant: "Throughout the whole game, no time is deducted if the player plays in less than one minute. The Japanese expression for this is literally "no time" (in katakana). So when a player is said to have played in no time, it actually means within fifty-nine seconds rather than instantly."

Not sure if that applies to this match, though.

Some digital chess clocks have a byoyomi setting. Probably called 'countdown' or 'byoyomi' or something.

[MartinCarpenter]

I do not want to minimize the recent achievements with go, but I'm not sure whether those achievements are based on a revolutionary approach or just the standard google approach to throw a lot of money and a LOT of computing power onto a problem.

You can read the papers and things - there were links to it in the articles after they beat the European champ. iirc on Chessbase. What they're doing really is very, very different to how computers play chess.

It still takes monstrous computation power of course (esp if they're going to have a hope in this match!), but its nearly all spent in the evaluation routine(s). Neural nets driven by distributed GPU computation. Searches vastly fewer positions than chess programs and even fewer than previous Go programs.

Chess simply is, when you get down to fundamentals, a brute force friendly problem. Dumb brute force + modern CPUs basically solves it in terms of playing very good chess. No to belittle it for one moment - its been a huge long term effort! - but the rest is all 'engineering'

It would actually be fascinating to see:
a) How strong could an 'evaluation first' chess engine like this Go engine actually get? A bit like a very strong human who never badly blunders. Probably never as strong as the brute force searchers, but the contrast would be fascinating.

b) Building the strongest correspondence chess player in the world - that would I think be a reasonably straight forward process using these sorts of deep learning techniques to control the search results from brute force engines.

Those would both take quite a bit of actual effort to make them work, and the pay off is small, so you can't imagine it happening. Go is huge publicity (esp in China/Japan etc), so obvious why they're doing it.

[Brian Towers]

I was probably insulting the stronger players I was playing by playing on in those situations.

I don't think so. It's more of a faux pas. You're just reducing your score.

I think I would also pass before that point was reached, which would prompt polite constructive criticism after the game. I was also never able to actually keep track of how much territory I had and whether I was ahead or behind overall. Bit of a handicap that!

It's not the counting bit which is tricky. It's the question of life or death. If you are stopping too early then it's because you've incorrectly assessed life/death. One of your groups is in limbo. Adding one or two stones could cement life and add territory, for example. Or there is a ko fight you must win.

Is there an etiquette in Go regarding passing - such that you shouldn't try and bluff around things such as passing? And again, I wonder how that is handled with computers. Is there a 'handler' for AlphaGo that will pass or resign on behalf of the computer?

Passing isn't really a question of bluffing. It is either right or wrong and this is something that computers are particularly good at, a bit like tactical combinations.

Here's a question: if the board in Go was expanded beyond the current 19x19, to say 23x23, would that favour humans over computers again, or not?

That's a tricky one. A lot depends on how the computers are programmed. It's clearly not just brute strength calculation although clearly that's useful for fighting and life/death. I don't know the extent to which they rely on opening books of joseki (corner openings/fragments) and fuseki (whole board openings). I suspect a lot less than in chess. I think this is one of the things that makes go a more difficult game than chess. Fuseki and joseki are more guidelines. Chess openings are more absolute. It's difficult to explain if you haven't played both games.

The problem with enlarging the board is that it changes the balance between the importance of corner/edge territory and centre influence. One way to think about this is make a line of stones along the 3 line. The owner of those stones owns the edge territory. Any stones the opponent places inside are dead. Suppose while the player was building the 3 line his opponent was using his stones to build the 4 line. The 4 line doesn't give him control of the rest of the board but it does give him enormous influence on what is the much larger territory and it may be an even trade, small territory v large influence. When you expand the board you change the ratios between those two areas and so you change the wisdom of those strategical extremes and the thinking in general about space. How will the human cope with that and how easy is it to change the computer's "perception"? I don't know but I suspect the computers can have their programs suitably modified (which obviously also involves some extensions of theory) quicker than the humans can modify their intuition and feel.

At what point does expanding the board become silly

When it takes too long to play a game!
You could think of 9x9 and 13x13 as the go equivalents of blitz and rapid

[MartinCarpenter]

The current computer in question basically plays by 'feel' anyway, so.....
(Given that a major part of its feel is based on games played against itself it might well update that faster than a human could!).

[Paolo Casaschi]

This bit may also be relevant: "Throughout the whole game, no time is deducted if the player plays in less than one minute. The Japanese expression for this is literally "no time" (in katakana). So when a player is said to have played in no time, it actually means within fifty-nine seconds rather than instantly."

In other words, I only start using the "initial" 2h time allotment after I exceed three times the 60s per move limit. Correct?
Couple of questions for personal curiosity:
- how many moves is a typical go game?
- is zeitnot and losing on time more or less of a problem than in chess?

[Roger de Coverly]

In other words, I only start using the "initial" 2h time allotment after I exceed three times the 60s per move limit. Correct?

In chess terms I believe it's as if you had G/120 and when that expires, you get 60 seconds a move, with a feature that you are allowed two "losses on time". Is the 60 seconds cumulative or delay?