Have we been duped by AlphaGo?

[Kirby]

Yes, it is contentious to me - perhaps more, because of the Disney talk, but contentious, nonetheless. Just look at the quality of my Malkovich games for an example. I spend much more time on that kind of a game, but feel worse about the quality.

I can't prove anything about the quality like some of you seem to claim.

So I disagree with the opinion here - so what?

A bigger problem is that I can't help but express it, and this thread continues.

[Kirby]

To further clarify my opinion, I personally suspect that there is some range of time limit for which increased time helps produce a better quality game. But saying that that 6 hours is better than 3 hours is arbitrary. Ok, maybe you analyzed the commentary from these games. Then your decision of game quality is based on commentary and not time limit.

[jeromie]

This may have been referenced elsewhere on the forum recently, but it seems appropriate given Kirby's introduction to his latest comment: http://xkcd.com/386/

Interestingly, the debate about how time limits affect a player's blunder rate is something that computer analysis can probably help us answer objectively. If we were to take a significant sample of a player's games at different time limits and run them through a program like CrazyStone, we could get a good idea about how time controls affects performance. My (non data informed) guess is that the blunder rate is measurably lower at higher time limits (up to a point - at some level additional time won't make any more difference) and that there is less variation in a player's performance from one game to the next when playing with higher time controls. I also think the impact will be most noticeable at higher levels of play. As a 5k, any variation based on time controls is probably overwhelmed by my general propensity for mistakes unless we are talking about extreme settings (i.e. Ultra blitz).

We don't yet have access to a tool that can accurately analyze professional games, but I'm looking forward to when we do. Humans will still have to explain why a move was a mistake for the foreseeable future, but this will at least mitigate the impact of differences in commentary style and/or quality. Then we might be able to actually determine when additional time stops making a measurable difference in human performance.

[xed_over]

let's think about this logically...

With each move, there are a number of possibilities to consider, and each of them have a number of permutations to think about. The more time available, the better likelihood of picking a move that leads to higher winning percentages. The less time available to consider all the options, the more likely to pick a bad move.

How much more time? How much less time? That's not really the point. Its more of a generalized statement that more time can lead to better games, while shorter time limits can lead to more mistakes.

[John Fairbairn]

Just look at the quality of my Malkovich games for an example. I spend much more time on that kind of a game, but feel worse about the quality.

I can believe that, but you are an amateur. I vividly remember two shogi pros invited to tea here in London (i.e. well away from the usual sorts of places where they'd fear being overheard) saying that "thinking time is dreaming time for amateurs" and then falling about laughing so much one of them nearly fell off the sofa.

Their point was that a pro can make good use of extra time - he knows what to think about. An amateur, even without a blindfold, is just trying to pin the tail on the donkey. He lacks the conceptual tools to go much beyond that strategically. In a game an amateur needs just enough extra time to avoid gross blunders (such as missing snapbacks) and maybe do a bit of counting. Beyond that: dream on. Amateurs should spend extra time on study instead.

I have yet to come across a pro who does not believe he could play better with more time. There are, to be sure, some pros who prefer to play with fast time limits because they believe they are more likely to win prize money that way, but that is not at all the same as saying they thus improve the quality of their play.

When the Nihon Ki-in was formed, time limits were a BIG issue. But for reasons unconnected with quality of play. One of the main concerns was the physical effects on young players, and so they typically had 8 hours each instead of 13 hours (Suzuki Tamejiro famously insisted on 16 hours each, yet still lost a game on time). The post-war hardships (especially low-calorie rations) put strong downward pressure on time limits, and later the booming sponsorship of go in Japan led to a demand for more games, and so time limits eased yet again to make time for these. The continued trend downwards can likewise be linked with commercial pressures and changes in fans' interests and attention spans. I have yet to come across a comment that suggests time limits are lower because pros asked for that - as far as I can recall (it would certainly be rare nonetheless).

Extra time is associated with higher quality in another sense. The most prestigious title matches span seven games, and generally the more final games there are the more the title is esteemed. The popularity of repechage tournaments can also be included under this heading. There is a general feeling that no-one wants to see the best player drop out early because of a single bad mistake. More games=time enables his quality to emerge.

[Kirby]

How much more time? How much less time? That's not really the point. Its more of a generalized statement that more time can lead to better games, while shorter time limits can lead to more mistakes.

No, it is exactly the point. We are comparing games played under specific time limits, so the argument is whether that specific time difference results in a different quality of game.

The generalized statement is what I disagree with. Specifics matter when making a conclusion about a specific situation.

[Kirby]

I have yet to come across a pro who does not believe he could play better with more time. There are, to be sure, some pros who prefer to play with fast time limits because they believe they are more likely to win prize money that way, but that is not at all the same as saying they thus improve the quality of their play.

FWIW, at the US Go Congress - I think in Tacoma a few years ago, Takemiya indicated that the time limits for the master's tournament were too long, and that they should be shorter. Players in the master's tournament aren't all professionals, but the level is very high - and professionals do participate. And I don't know Takemiya's precise reasoning. But clearly, there are some pros that find shorter time limits to be better in some circumstances, outside of simply winning prize money.

Maybe he thinks longer time limits are important for professionals - or maybe he thinks he can play better with longer time limits.

I don't know.

But the idea that longer time limits equate to better quality, *universally*, I really believe is bogus. I recall a book, "The Art of Learning", where the guy is a very good chess player. He looks back at his games and finds that his "brilliant" moves were moves where he spent sufficient time, but not too long of time.

To be sure, pros are better at managing longer time periods than amateurs. But they are better at managing shorter time periods, too.

[Bill Spight]

Just look at the quality of my Malkovich games for an example. I spend much more time on that kind of a game, but feel worse about the quality.

It is possible to overthink plays. However, given the pace of Malkovich games, there is time to overcome temporary delusions. Perhaps you feel worse about the quality of your Malkovich games because you appreciate the difficulty of the game better and realize how little you understand it. (That's true for all of us, OC. ) By contrast, you may be less aware of your mistakes and doubtful plays in everyday games.

[Bill Spight]

Interestingly, the debate about how time limits affect a player's blunder rate is something that computer analysis can probably help us answer objectively. If we were to take a significant sample of a player's games at different time limits and run them through a program like CrazyStone, we could get a good idea about how time controls affects performance. My (non data informed) guess is that the blunder rate is measurably lower at higher time limits (up to a point - at some level additional time won't make any more difference)

A related question is how long players take on blunders by comparison with good plays. OC, careless play produces blunders and is quick. However, once careless play is avoided, I am not at all sure that there is a relation between blunders and the time taken to play them.

Back when I was competing seriously, I followed a suggestion of Botvinnik's and made note of plays on which I took a lot of time -- more than one minute, I decided. Botvinnik's point was that the positions that you spend a lot of time thinking on are positions that you find difficult, and deserve study, no matter what your eventual play or how good or bad it was. Well, positions that you find difficult are the ones where you are most likely to blunder. Both slow plays and fast plays can be blunders, for different reasons.

[Anzu]

Takemiya Masaki once spent .. five hours on one move..

I think the longer the time limits, the better the play.

[Kirby]

Perhaps you feel worse about the quality of your Malkovich games because you appreciate the difficulty of the game better and realize how little you understand it. (That's true for all of us, OC. ) By contrast, you may be less aware of your mistakes and doubtful plays in everyday games.

It's a good point, and it's very possible. I also admit that it's possible that there is a higher quality in 6 hour games vs. 3 hour games, for example. But I think it's not necessarily the case, and the conclusion is a little arbitrary. The only evidence we have toward that is the comment suggesting that more pro commentary finds more mistakes in 2 to 3 hour games. I'm also open to the possibility that 3 hour games result in higher quality than 6 hour games. Three hours is still a lot of time, and I think it's possible.

Someone on the forum sent me a PM saying that life was not worth arguing about this, and that I should just drop it. Strictly speaking about quality of play for 6 hour games vs. 3 hour games, maybe I agree with him.

But one of the reasons I keep responding to this thread (other than my stubbornness) is that I get a little angry and worked up with the tone regarding modern time limits. Calling them "Mickey Mouse" time limits, and "full of mistakes", and things like that seem to belittle something that I respect.

And I think that is worth arguing about, even if it ruins my weekend.

[erislover]

I don't know where I sit exactly, but I think I would agree with Kirby. First, even if we accept that slower games are better, that doesn't imply that slower games change results—the better player will win fast or slow, we'll just hear more pro criticism of the moves in fast play. And I don't know that this must imply that the games leave us little to learn because of a higher mistake rate; what assumptions must be made in order to say that the mistakes made in fast play necessarily mask novel, instructive play? —I can't even begin to formulate the deduction chain here.

[Charles Matthews]

As far as play is concerned, there is little in go that is novel.

Isn't there a quote from Sakata saying that "technique" is so much improved, since the time of (say) Dosaku? We're in danger of misunderstanding what pros mean by technique, of course. But e.g. standard techniques for enclosures do have a history, 4-4 point strategies have a history.

As far as the OP's view goes: when I was a mathematician, in fact at my first conference, another mathematician told me that when you "know how it works", i.e. have solved a problem, there comes a moment of disappointment: "just that!" The AlphaGo phenomenon, to use a fairly neutral term, does suggest that deep mysteries of the middle game may be amenable to comprehensible simplification techniques, at top pro level.

"Middle game fighting" is not perspicuous, in human perception, in great games. Perhaps styles that lead to earlier shape-fixing (to come back to Sakata) help to make it so, and perhaps players with more elusive styles might have more chance against AI using current techniques. If not, the slightly let-down feeling might have some justification.

[Bill Spight]

As far as play is concerned, there is little in go that is novel.

Isn't there a quote from Sakata saying that "technique" is so much improved, since the time of (say) Dosaku? We're in danger of misunderstanding what pros mean by technique, of course. But e.g. standard techniques for enclosures do have a history, 4-4 point strategies have a history.

I was intending to contrast plays with concepts. As I said:

AlphaGo is unable to articulate any new concepts, and so what we can learn from it is limited.

Players come up with new ideas in go, and IMO now is a particularly creative time. But all we have from AlphaGo and other strong computer programs are plays.

Now perhaps AlphaGo will come up with a new play in a joseki, or a new joseki. That would be a new idea, I think. But suppose that AlphaGo makes a play that surprises the pros, and that seems to be good, or at least OK. Any new concept that arises from that play depends upon human analysis and creativity, as well. What does the play accomplish, and what features of the position are relevant? Humans need to figure that out.

[Charles Matthews]

As far as play is concerned, there is little in go that is novel.

Isn't there a quote from Sakata saying that "technique" is so much improved, since the time of (say) Dosaku? We're in danger of misunderstanding what pros mean by technique, of course. But e.g. standard techniques for enclosures do have a history, 4-4 point strategies have a history.

I was intending to contrast plays with concepts. As I said:

AlphaGo is unable to articulate any new concepts, and so what we can learn from it is limited.

Well, perhaps I misunderstood what you were saying. However, I think do I agree with the premise A (AlphaGo is unable to articulate any new concepts); but I don't agree with the conclusion B (what we can learn from it is limited).

Players come up with new ideas in go, and IMO now is a particularly creative time. But all we have from AlphaGo and other strong computer programs are plays.

I also have a problem with this comment. We have, in the field of computer go and human pro go, game records, and commentaries on them by pros (and go writers strong enough to write journalism and books about high-level games, who are presumably at least 6d ama).

For ordinary pro go, we in addition have such commentary and expository works about the games that are written by the participants in the game.

That is the difference; and while it is obviously a plus to have the participant view, I don't think it is as disabling as you imply. We cannot in any reasonable sense have the "intention" of a go AI to discuss; but we could at least in principle, from a stable version of a go AI, reproduce the considerations that led it to play the way it did.

Now perhaps AlphaGo will come up with a new play in a joseki, or a new joseki. That would be a new idea, I think. But suppose that AlphaGo makes a play that surprises the pros, and that seems to be good, or at least OK. Any new concept that arises from that play depends upon human analysis and creativity, as well. What does the play accomplish, and what features of the position are relevant? Humans need to figure that out.

Suppose Kitani came up with a novelty of this kind (happened often); while it was left to Kitani disciples to articulate just those things. We could either take the line that the "Kitani tradition" is doing the exposition you want; or we could be a bit more sceptical about that reification, and say that some writer is claiming authority for what are personal views.

I think the line you are taking is a bit of a stretch, really. There is a connection between concept formation and the need to write commentaries (i.e. to evolve a jargon adequate to describing pro go at some level such as "good amateur"). But there are other ways to improve perception in go, or at least it appears so to me. In a typical multiple-choice question "Black to play at A, B, C, D, E" we do want more than "after A Black is better", but adding a short plausible continuation can show that to the satisfaction of some class of readers that "after A and the continuation 2 to 5, Black is better", as long as we have a way of generating continuations. Which software such as we are discussing can handle.

I don't endorse a pessimistic line on such matters, when it hinges on the "incomprehensbility" of computer go. For games in general, it appears to be a contingent matter: game G might be such that its gameplay made human search of variations that are not very long considerably weaker than exhaustive machine search. This can be called "unfairness" of G; in the sense that go players might understand in terms as "in this position, the bad shape move A works well, a kind of blind spot".

Both AlphaGo and Lee played candidates for such moves in the match, but I don't think the perception is of a lack of "equity" in the gameplay.