I just noticed this interesting post from reddit/baduk: https://www.reddit.com/r/baduk/comments ... e_provsai/

I wonder if the assertion that top pros will never need to take 4 stones handicap from AI is correct.

_________________
Sorin - 361points.com

It is only three years ago that no one believed that AI's would compete with pros in our life time. Meanwhile it is clear that AI's are significantly better than people at our beloved game and that the gap is widening (really the point of the article). If Shin Jinseo is correct and AI's will never be able to give four stones, regardless of how much better they become, then we are talking about the limitations of Go itself (played on 19X19 boards, with existing rules). At that point I think the whole issue losses the tag "interesting".

But don't forget...

"Forecasting: To observe that which has passed, and guess it will happen again."
- Elbert Hubbard (1859-1915)

So there is little reason to trust Shin Jinseo's assessment at this point in time. Personally I expect that he is being a little too optimistic about the human side, but only time will tell!

_________________
Dave Sigaty
"Short-lived are both the praiser and the praised, and rememberer and the remembered..."
- Marcus Aurelius; Meditations, VIII 21

( 4 years ago, at least Deep Mind already knew they were close. )

Jan 30, 2016: beyond top humans People can be slow learners, can't they.

I expect an AI which is trained to play handicap games vs pros and ellicit their typical mistakes, overplay in ways it knows they can't punish, use the psychological tricks of handicap go etc would be able give more stones to a pro and win than a perfect player could, and that this number is >=4. So maybe pros are 4 stones from God and 6 stones from the Devil.

.

Our lifetime varies for different forum members, but Zen was starting to beat EGF 6ds iirc just before AlphaGo came along. I made a bet with Natasha Regan (chess and go playing friend and co-author of book on AlphaZero chess) that computers wouldn't beat a top pro in match conditions within 5 years. That was prompted by a Facebook discussion around GM Nigel Short basically saying Go wasn't so hard and if only someone threw the resources at it it would be cracked. I thought he was ignorant of the efforts that had gone into computer go (which weren't on IBM Deep blue level but more than man and dog in shed) but he was kinda right. I ended up losing that bet because Demis Hassabis learnt go at Cambridge, founded DeepMind and got bought by Google so was well funded ( and the happy coincidence that neural network techniques being refined for image recognition applied well to go, policy networks predated AG but value network was their breakthrough). I wouldn't have made that bet about the weak end of pros. Beating Fan Hui was ahead of schedule, and then the speed of improvement to beating Lee Sedol was impressive (I think more so to go players who appreciated the level difference between them than people in AI community) but when that happened I remember people saying this was 10 years faster than expected rather than 50 (my expected lifetime) or 100.

Who knows what the future holds? But as for today's top bots, running on today's hardware, I think it likely that humans will come very close to their level of play. The reason is that the main difference between them and humans lies in the realm of evaluation, what we call intuition, and what humans are able to codify as go knowledge. They do whole board reading better than humans, but humans can and do surpass them in local reading in certain types of positions. (I expect that go playing programs of the future will advance in that regard, because developers will program them differently.)

We humans are quite clever, and quite good at learning things that we cannot articulate. We are also good at coming up with useful concepts and rules of thumb. My guess is that pros 50 years from now will be 2-3 stones stronger than today's pros. And that average amateurs will advance even more than that, because everyone will have programs to review their games and answer some of their questions.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

That's a good guess, Bill; but not if Skynet wins first.

Why do you think I keep helping training Leela? I want to be in its good graces when the time comes...

Take care.

_________________
玄 之 玄

I doubt trickery would worth 2 stones at pro level (unlike ama levels).

I doubt this too. Even before the bot era, I had the impression that at top levels games are decided 99% by reading. Just like in chess, the difference between champions and other 9ps were a halfmove or so reading depth. And I don't think humans can read significantly more than today, so if - say - a stone can be gained that must come from better intuition/evaluation. There definitely is some room there (study with bots, patch blind spots), but unlikely 2-3 stones.

People with weak hardware can notice how bot evaluation quickly changes by significant amount at the beginning of the search. This is the transition from evaluating the position to reading the position out. Bot evaluation, in my opinion, may be even weaker that human. For example, LZ is doing NN evaluation by randomizing board rotation each time. And it was reported in complicated positions its evaluation can swing by 50+% between different rotations (even though they all received the same amount of training). So its accuracy cannot be too good.

But it doesn't really hurt since what matters most is accurate reading. Not in the minimax sense, but Go apparently does not need that - reasonably good candidates are enough usually (esp if one avoids all-in positions, sharp tactical lines).

About the original assertion: taking 4 stones with komi only means 3 stones strength difference. And bots still likely have significant room for improvement.

I think perhaps the initial statement might be being misunderstood.

Assume there is such a thing as "perfect play". Neither the best AIs nor best human pros are there yet, so we could talk about "by how many stones are they weaker than perfect play". Now we don't KNOW the answer to that, but suppose the best human pros are now three stones weaker than perfect play. Do you not see that "the best human players will never need four stones against the best AI" (no matter how much the programs improve or machines become more powerful) would necessarily be true.

Mike: Agreed that’s important—it’s not necessarily arrogance, or downplaying AI. Though like Uberdude said, overplays complicate the question.

Btw: I think that a few years before AlphaGo, there was a poll on here about when AI would reach pro strength. Iirc, the most common result was 10-20 years. So that was overly pessimistic, but shows that people thought strong AIs were on their their way.

I’m on my phone and failed to find the thread.

_________________
Occupy Babel!

This is the assumption being debated. 3 stones seems underestimate, considering there are already 2 stones to strong bots, which are still in improvement. Bot strength also depends on hardware, competition games are run on orders of magnitude stronger hw than a good gpu at home (and hw is also in improvement). I also don't think NN bots can get closer than a stone (or a half stone?) to perfect play, they also have their intrinsic weaknesses.

It may not effect pros as strongly, but there is some sort of psychological aspect in handicap go.

I heard from a member of this forum awhile back that the appropriate handicap between two players is around what the weaker player thinks it is. Maybe I’m not saying this in a precise way, but the idea is like this:

Suppose that, based on difference in technical skill, there should be a 3-stone handicap between two players. Now also suppose that the weaker player knows that his opponent is stronger than him, estimating about a 4-stone difference in strength. Even though the “real” difference is 3-stones, since black is psychologically beat, there’s a good chance white can win at 4-stones.

Not sure how accurate this is, but go can become a harder game when you lack the confidence to beat your opponent.

_________________
be immersed

The point has been made: humans can't play better than the best AI bots. Likely, research in developing a strong AI bot will dwindle; What's the point? Probably human-played go could make use of AI bots the way the chess population does, using the bots to analyze games. We already do that but not on a large scale way. I'm thinking of how large on line chess sites (chess.com) have built in bot analysis available right after your game finishes. No need for expensive graphic processors or expensive access to cloud based virtual computers.

Well, reading is a complex skill involving calculation of variations, candidate move selection, and evaluation. It may well be that at go, top level games are largely decided by depth of calculation. But that's largely because the players share assumptions about move selection and evaluation. That's one reason why you see flipflops in bot evaluation of pro games. One pro makes a mistake and their opponent immediately makes a compensating mistake. Without the shared assumptions the opponent might easily be able to profit from the first mistake.

Chess makes a good comparison. It has long been believed that chess is mainly tactical. The figure, going back at least to the early 20th century, is that chess is 98% tactics. Back in the early days of computer chess, programmers tried to emulate human thinking, which is not very tactical by comparison with even the ability of computers of the time. The breakthrough in computer chess programming came from switching to calculating variations instead. (OC, things are not that simple. ) Computers still needed opening books (chess knowledge) to play well. The result is that, before todays neural network programs, computer play came to look quite different from human play. That difference is no small part of detecting cheating by using a computer program to select moves.

By contrast, the tactical approach to computer programming did not produce strong go programs. The final breakthrough which did produce superhuman strength came from the use of neural networks, at least metaphorically an emulation of human brains and our unconscious parallel processing. Yes, we can tell differences between human and computer play. Computers pincer less, like 3-3 invasions, play attachments and shoulder hits, tenuki often, value corners more and sides less, value moyos less, but still like central influence, make 2d line plays more, play the center differently and better, and so on. Except for center play, all of these differences are easily learnable by humans. And as for go knowledge, within a few years it will be possible to produce new joseki books based upon computer play. In chess the analogy would be to rewriting opening books.

Neural network chess engines now play at the top level. Yet they are certainly weaker at the calculation of variations than other engines, and their play looks more human. I expect that that will make it easier for humans to learn from them than from the previous generation of tactical engines.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Bill: not at all disagreeing with the bolded bit, but maybe putting a gloss on it...

Japanese pros have been trying to package some of the new AI concepts by putting a label on them - standard human practice, of course! Whether any of the new terms will stick we will have to wait and see, although "direct 3-3" (using the English word) for the early 3-3 is one that may last as I think there's a new book about it. As an aside, at first I thought that was a bit of an iffy term, as I couldn't see what was wrong with the more obvious "early 3-3", but I've come to appreciate that it captures in a wee way the bull-charging AI style of not bothering about keeping options open.

But the main term, which I liked and want to mention here, was "network." This seems to be the new black for moyos. Obviously it has a touch of elegance in alluding to value and policy and neural networks, but I think it also captures the new emphasis on "associations" between stones rather than "connections." Here is what I believe is a good example:



The triangled last move completes, I infer, a network. This was by Shibano Toramaru, one of the new AI-focused generation of course, against Akiyama Jiro, an older player perhaps not quite as AI-focused. It was from the NHK Cup game published this week. Already White I some way behind. Both players kept pace with Lizzie until just before the triangled move when play was in the lower-left corner. There, Black left White a little overconcentrated - early overconcentration seems to be one of the main strategic goals of the bots, according to Japanese pros. And note how the Black network is the very antithesis of overconcentration.

I haven't seen anything that explains how you would define a network (though early high forcing moves seem to be part of the toolbox), let alone how to play with or against it. But what seems to happen is that you get more territory bang per buck. In the old moyo paradigm what you were mapping out was virtual territory and you assumed the opponent would invade so that you ended up getting that of the virtual territory as real territory.

With a network, invasion seems to lead to getting more like two-thirds of the prospective territory. Invasion is therefore replaced by reducing moves, even high ones.

All these high moves, both for building and for erasing the network, seem to have "associations" (my term) emphasised by constant tenukis. I hae further have a feeling that AI bots dislike pincers because they interfere too much with association-building. I also believe associations has a wider connotation than aji.

Obviously this is mostly speculation by me, an amateur, but at the very least I feel sure that finding the right terminology is going to be a big step forward.

I spy a nozomi at f5, another favourite of AI-style

We are all feeling our way, as mere humans.

New concepts will emerge. I like direct 3-3, BTW, as it can apply to the 5-3 and 5-4 as well as the 4-4. IMO we need good terminology, good definition or description, and a good set of instances, probably a good prototype. Easier said than done.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

I wonder how did the white shape C4,D4,E4 occur. It seems a bit over consentrated.

Matti, here's my guess for move order:

Click Here To Show Diagram Code

[go]$$Bc
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . 7 . 0 . . . . . 9 . . . . . |
$$ | . . . 4 . . . . . , . . . . . 1 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . 8 . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . 5 . . |
$$ | . . . 2 . . . . . , . . . . . , . . . |
$$ | . . . . . . 6 . . . . . . . . 3 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

Not sure if 13 or 15 would come first, but probably this way as submarine against the big knight is unusual as gives bigger corner to white than normal. But with white already defending with 14 c4 is a tad soft: white could play from outside and let black live in corner but then f5 is probably a good exchange, so in a fast TV game white probably went for solid cash,

Click Here To Show Diagram Code

[go]$$Bcm11
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . X . O . . . . . X . . . . . |
$$ | . . . O . . . . . , 2 . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . 1 . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . 8 . . . . . . . . . . . . . . . . |
$$ | . . . . 9 . . . . . . . . . . . . . . |
$$ | . . 7 . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . 0 . . . 3 . . . . . . . . . . X . . |
$$ | . 5 6 O 4 . . . . , . . . . . , . . . |
$$ | . . . . . . O . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

Click Here To Show Diagram Code

[go]$$Bm21
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . 2 X . O . . . . . X . . . . . |
$$ | . . . O . . . . . , O . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . 1 . . . . . , . . . . . , . . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . X . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . O . . . X . . . . . . . . . . X . . |
$$ | . X O O O . . . . , 3 . . . . , . . . |
$$ | . . . . . . O . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

I like the term "association". I wonder whether the traditional description of stones "working" with other stones is related.