Have we been duped by AlphaGo?

[bayu]

Unfortunately the write-up stopped after game 2. I really liked the commentary by Zhe Li (6p) here https://massgoblog.wordpress.com/2016/0 ... lee-sedol/, which was also referenced in a thread on lifein19x19.

Concerning duping, the time settings are not that important in my view. Humans and computers benefit differently from longer time limits, but both benefit. Maybe Yi Se-tol would have won a game more with other settings. In half a year, the result would be a clean 5-0 anyway, independent of settings. In the commentaries about the AlphaGo-Yi matches I saw, I didn't spot any critical ??s due to time issues. Well, AlphaGos blunder in Game 4 could be considered one, which probably wouldn't have happened with more careful analysis. So, the question is, how many !!s did we miss. Yi pulled off a fair share of them and toyed with AlphaGo (see above link). It was certainly magic for me. Unfortunately I am too young to have seen the good old days (and I am extremely grateful for JF to share them). I'd be very happy to see a longer time limit game with AlphaGo. It might be the computer that surprises us, not the human.

More important concerning duping is the momentum of surprise. AlphaGo was fed all games by Yi Se-tol, while he only had the games between a considerably weaker AlphaGo and Fan Hui as study material. Showing Yi Se-tol some game records of AlphaGo versus itself would have had an impact. Or allowing Yi to play a test match up to move 150 which doesnt count for the match. These things are never done, though, for a variety of reasons. The next person to play AlphaGo has at least the advantage to be able to study 5 games between a considerably weaker AlphaGo and Yi Se-tol.

[Javaness2]

It was said that Kasparov was the wrong choice to defeat Deeper Blue because he wasn't a positionally based player, he was more of the calculator. Is the same going to be said of Lee Sedol? Of course, if you want to generate a discussion, you can certainly advance that opinion, but it doesn't really have the same weight behind it. If you transport Go Seigen back to the Edo period and sit him down to play AlphaGo he might win this month, but probably not next month. So where is the hood and who is that winking? My own expectation is that the next pro to try their luck won't win a single game.

[gowan]

First, can AlphaGo teach us new things about the fuseki?

AlphaGo doesn't make any distinction between what we see as different stages of the game.
Maybe this is something to learn - that there is no such thing as "fuseki", it is just about doing one's best from beginning to end to try to win.

Things like opening, middlegame, endgame, fuseki, etc., are human constructs to help us understand and play the game. Obviously they have no meaning for AlphaGo since AlphaGo doesn't "understand" concepts or really understand anything the way we do. Nor does it appreciate the beauty and thrill of playing the way we do.

[Javaness2]

I'm not sure why 5 games should provide a solid basis to redefine our view of fuseki.

[erislover]

I have a hard time knowing how much of this OP is tongue-in-cheek and how much is real opinion with a reduced-fat whipped cream topping. I don't know why we should look to any particular era's fashion in go openings and consider them canonical. Why should 3-4 points be preferred over 4-4 points, just because Japan used them for so long and China didn't? Shin Fuseki showed that there was a lot of unexplored space in go which was held back only by tradition. Pros still play "weird" moves—though even using such a phrase belies some "weird" biases. Frankly, I think if move one was a 3-4 point then under perfect play move one lost the game, and I fully expect that when I pass on and meet the great go player in the sky all of my prejudices on this matter will be confirmed: the best first move for black is obviously the 4-4 point (specifically Q4 though the reasons for this over Q16 are quite technical and the space here is to small to adequately explain). The fact that alphago used a 3-4 point proves it still has a ways to go. Nevertheless, for a komoku player, alphago was not bad at all.

[Mike Novack]

I think there is a certain amount of misunderstanding here, mostly related to what a neural net is and how it is trained to be able to do anything (and on what, in particular, was AlphaGo trained).

The PROGRAM part (of a neural net) is code that simulates the neural net. The only go related part of that is that the "geometries" of the neural net might be particularly suited to efficiently be trained to go << for example, take advantage of the known symmetries, the dimensions of the board, etc. >>

Once it exists, the neural net is TRAINED to the task. In this case, a very large database of high level games (or games at a lower level if that's the desired level of play). In other words, if the program doesn't make the move the expert (human) chose in that situation, cell values are given a "kick" until by trial and error it does. Of course a different human expert might have chosen a different move or the situation very slightly different in a different game. The training data is huge.

So "learn something new that humans didn't know before?" Well they DID know it in some sort of collective sense. Not quite accurate, but the way you might imagine a neural net like AlphaGo working is that inside were a largish group of human experts coming to a consensus about what move should be chosen. When we see AlphaGo make some novel move that was derived from move HUMAN experts played in a large number of (slightly) different situations. We humans may not be good at looking at a set of different data and extrapolating the correct response in a situation slightly different from any of them, but a neural net is.

[erislover]

Mike Novack, AlphaGo reached its strength through self-play. The database games were basically used as a foothold for self-play training. AlphaGo shows us that the proverb should be "lose your first ten million games as soon as possible."

[Polama]

At the professional level, do longer time settings produce generally better games (fewer blunders, more brilliant moves)? Yes. If not, I'd expect the players to use less of their time when they have it.

At the professional level, do longer time settings produce better players (if you compete exclusively at 2 day matches will you reach deeper understanding)? Unclear. We can all have our hypotheses, but either is plausible and an experiment is unlikely to be performed to answer this.

At the amateur level, is it better to study fast or slow games? Unclear. This could at least be studied, but I don't see it happening. On the one hand a longer game has more brilliant moves. On the other hand, those brilliant moves usually make use of the extremely specific conditions of that game. In a shorter game, I expect professionals make more use of good shape and common patterns instead of deep search for the truly best move. An amateur might still learn good shape from an instance that turned out to be a blunder. Either hypothesis could be true.

Was Lee Sedol the best choice to go against AlphaGo? From statistics, probably not. Whatever weaknesses AlphaGo has, some other top professional is likely to have happened to have matched up slightly better. Without playing matches against them all, though, I don't think you can state who that is though.

Does it matter? Not really. It looks like Lee Sedol played at a top professional level in the games. We'd be talking fractions of a stone difference. If somebody could have edged out AlphaGo, it would have been a temporary triumph anyways. Also, if somebody plays well against this particular algorithm, that doesn't imply that their style is superior in general, just against this algorithm.

[hyperpape]

First, can AlphaGo teach us new things about the fuseki?

AlphaGo doesn't make any distinction between what we see as different stages of the game.
Maybe this is something to learn - that there is no such thing as "fuseki", it is just about doing one's best from beginning to end to try to win.

Things like opening, middlegame, endgame, fuseki, etc., are human constructs to help us understand and play the game. Obviously they have no meaning for AlphaGo since AlphaGo doesn't "understand" concepts or really understand anything the way we do. Nor does it appreciate the beauty and thrill of playing the way we do.

It seems to me like these are misleading claims. It's true that alphago does not have a stream of consciousness in which it thinks "well, it's the fuseki now". More than that, it may be that there's nothing you can point to in the neural net and say "fuseki lives here" (but can you do that with the brain...?).

However, neither of those things justifies saying it has no meaning for AlphaGo. AlphaGo plays a fuseki. It does not immediately start fighting without preparatory moves, at least no more than trigger-happy modern players, so it plays a fuseki. The knowledge that makes it play that way is all somehow implicit in the neural network. But I don't see why that means we should say it isn't real. As far as it knows anything about go (and I'm not interested in arguing whether it really knows anything or blah blah blah...), it knows about the fuseki.

And if AlphaGo plays a fuseki, then either its play is novel and can teach us something about the fuseki, or it can't. So John's question is a good one.

[oren]

I was watching a short video about the top Shogi player vs top Computer player they had in Japan recently. They made each game a 2 day affair with each side having 8 hours of time. The result was the computer winning.

I think the AlphaGo time settings were fine.

[Bantari]

gowan, games with fast time limits make mistakes, but so do games with longer time limits. It is biased to point out only the former.

All humans make mistakes in all games. This is beside the point in this context.

What I know is that:
I make better moves when I have more time. I also make less mistakes. And the mistakes I make are less stupid. This is a fact.

What's more, this applies pretty much to everybody I know and ever played with. Both in Chess and Go. Over the past 30+ years of playing both games.

Do you have any indication or proof that the same principle that applies to us, mere mortals, somehow does not apply to the pros?

PS>
A simple thought experiment to support my (and John's) theory would be this:
Imagine you approach two top pros and suggested they play a match. One of them will have 2 hours for all his moves per game, while the other will have 4 hours per game. Do you think both would agree to this setup and consider it fair? Or do you think the one with 2 hours will think it not fair and will consider his chanes less than in even contest?

If you think the 2 hours pro has a good point to compain about this, you have your answer right there.
If we call it "Micky Mouse" os something else, the fact remains that it is very plausible to assume that longer thinking time produces better moves, on average.

_____
(*) Of course - there is some cut-off limit beyond which there is no difference or even decline - for example I would probably play better at 2 hours per player than at 16 hours per player per game. At 16 hours I would get tired, bored, and just resign. But if I took time to train for 16 hour games, who knows... On the other hand, I guarantee you that in 10 min sudden death I play much weaker than in 1 hour sudden death games. Unfortunately, so do my opponents - which supports my point.

[Kirby]

I should first note that I am replying again to this stupid thread because I have little self-control, and I am addicted to replying when I disagree with someone. It is not a wise decision, but it is Friday... And when have I ever been wise?

That being said...

All humans make mistakes in all games. This is beside the point in this context.

It's not beside the point, because the implication is that "more time" == "better quality game". I do not agree that this is always true.

What I know is that:
I make better moves when I have more time. I also make less mistakes. And the mistakes I make are less stupid. This is a fact.

1. You assume you make less mistakes. I find it hard to believe that you can conclude it as fact. For example, maybe you notice "obvious" mistakes during your review, but miss subtle ones that might be harder to find.

2. The assumption that increased time is linearly related to increased game quality is unfounded. Personally, I do *not* know it is a fact, but I also get the feeling that I play a better game if I have 30 or 40 minutes of time, compared to when I play a 10-second blitz game. But it does not follow from this that a 6 hour time limit will result in a better quality game than a 40 minute time limit. And in fact, I may overthink the situation and play worse.

Based on the last part of your comment, you seem to agree with #2, above.

Do you have any indication or proof that the same principle that applies to us, mere mortals, somehow does not apply to the pros?

No, and I don't think it applies to us, either. See #2, above.

If we call it "Micky Mouse"

Please don't. It makes me angry, and I think I have expressed this clearly by now (though, perhaps that encourages some people to do it more :-p).

[daal]

1. You assume you make less mistakes. I find it hard to believe that you can conclude it as fact. For example, maybe you notice "obvious" mistakes during your review, but miss subtle ones that might be harder to find.

2. The assumption that increased time is linearly related to increased game quality is unfounded. Personally, I do *not* know it is a fact, but I also get the feeling that I play a better game if I have 30 or 40 minutes of time, compared to when I play a 10-second blitz game. But it does not follow from this that a 6 hour time limit will result in a better quality game than a 40 minute time limit. And in fact, I may overthink the situation and play worse.

I do know for a fact that there are time settings where I make mistakes that I would not make in a slower game. Whether I make other mistakes is beside the point. Playing a bad move that I would not have played if I had had (and used) more time means that the game is worse than had I not played that move.

I don't think anyone said that there was a linear relationship between time and quality of a game. The question is: What is the relationship for the times in question. I think we all know that within the range of times that we typically play our games, that the quality of our moves is higher in slower games. It is also a fair point that given too much time, that there might even be a decrease is quality. This would imply that there is a time setting at which on a given day, a player would have the best chances to have the right amount of time to play his best moves. For me, this might be 60 minutes on a day where I've got nothing else to do.

How about for pros? Is their optimal time three hours or three days? Why have professional games gotten faster? Did the best players decide of their own accord that faster games are more suitable for modern people in a modern world than games lasting a day or more and that a highest quality game can be played in 3 hours, or was it that the world (sponsors) decided for them? Maybe it was a little of both. I could imagine that some pros felt that the benefit of an extra 5 hours brought diminishing returns as far as the quality of their games was concerned, and in the advent of televised games, I'm sure their sponsors also felt that way. Yet there may well be another fraction of pros that feel differently, that they can no longer play their best go and that the time settings are a constraint that reduces the chances of brilliant moves. I imagine that those people are angrier about the development of time settings than you should be about a hyperbole used to make a point. It seems to me that there is reasonable room for disagreement as to whether the current time settings used in professional games is optimal or not.

[Bill Spight]

And if AlphaGo plays a fuseki, then either its play is novel and can teach us something about the fuseki, or it can't. So John's question is a good one.

As far as play is concerned, there is little in go that is novel. Even the New Fuseki was not the first time that go players focused on the center. And at a more concrete level, the mini-Chinese formation goes back over 200 years. Contemporary professional go is more territory oriented than it was in the mid-twentieth century, but less territorial than it was in the eighteenth and nineteenth centuries. What AlphaGo has illustrated that center oriented, cosmic style plays are more potent than had been generally assumed. There is a lesson there, but AlphaGo is unable to articulate any new concepts, and so what we can learn from it is limited. If it plays enough games, perhaps some humans will be inspired to articulate novel concepts that describe many of its good plays. If so, that would be a real boon.

[Uberdude]

Kirby, I don't think anyone claimed more time will always lead to a better game (and I agree it won't) or that it was linearly proportional, I just think there is a (fairly strong) correlation. Is that really so contentious? When I play blitz I make many mistakes because I don't read enough and with just a few more seconds of thought I can see the mistake (sometimes lose to 1ds on KGS). When I play with 1 hour I play around 4d level, but when it gets to byo yomi I play worse, because I have less time! Of course I also sometimes make dumb mistakes in normal time, but I make much more in byo yomi. And when I have loads of time (hours or days per move) as I did when I played on OGS, I play at a higher level: when I was at BIBA in Korea and got my OGS games reviewed one of the 7d ama teachers said my games were 6/7d level, and I also beat breakfast 7d/3p there.
Anyway, I started a new thread to look at mistakes in pro games with different time limits: http://www.lifein19x19.com/forum/viewto ... 04#p204604.