Question for experienced players: Is this a thing? (Sideseki?)

General Go Discussion
45

You have me very confused here. How do you translate a probability of winning into points? And then how does that become “understandable”?

I think it’s definitely meant to be a territory estimation, but it’s only reliable in the endgame. And indeed if you look at it during the endgame, it’s usually a pretty good territory estimation.

If you really find it understandable in the opening, after only 8 moves of a 19x19 game, then you must have an incredible insight into the AI which I don’t have, and I’m interested into an explanation. But I have to warn you that I’m going to be extremely doubtful.

1 Like
46

In a game with komi 7, if the AI estimates the score as W+7, then it means that if komi is reset to 0 and the AI plays against itself starting from that position, then W has 50% chances of winning. It doesn’t mean the winning probabilities would be the same in a human vs human game.

1 Like
47

Afaik katago is explicitly trained to estimate the outcome of the game in points. This is a difference to for example Leela Zero. So I don’t think any translation is happening there.

5 Likes
48

I added level IV AI review to a couple of @JethOrensin’s games that applied this fuseki

6 Likes
49

Thank you very much for your help :slight_smile:

3 Likes
51

I don’t know: katago does that and I trust it. :sweat_smile:

I mean that “white +6.9 pts” is more understandable than “white 91.3 % winrate”.
I can easily understand an error that causes a few points loss, but I really can’t dig into that 91% that just sounds to me like “white won, forget about it!”

I wouldn’t say that.
But I think that the reference frame makes the difference.
If you look at that “+6.9” as a territory estimation of course you’re correct that it isn’t reliable after just few moves. That’s it.

But that isn’t how I understand it.
My starting point is: katago has no understanding about go. Not in the way us humans think.
It just has a black box that, for any given board configuration, comes out with a probability number. And that number IS reliable!
Then it’s up to us humans to understand what that number means.
The katago team managed to convert winrate into points difference and that’s the part I love most. Now it’s very easy to understand the weight of a single mistake: “you lost 6.9 points there”.

We can say that it’s reliable because when we sum up all those numbers in a completed game, the result is the actual score.
Well, it’s built with that goal in mind and that’s why it is consistent: because the training was meant specifically to sharpen that measure.

If you don’t think of it as territory but you read it as the value of a single move, all that comes together.

5 Likes
53

Indeed KataGo’s winrate and points (not a great name but it’s short and stuck) aren’t just different ways of looking at the same thing, like using celsius or fahrenheit or kelvin to measure temperate that can be easily converted between. It’s more like temperature vs mass, they are fundamentally different concepts (well, maybe not that different, as both related to winning the game). Indeed it is not uncommon to find them not increasing and decreasing in unison, e.g. one position might be winrate 70% and points +10, and then another might be winrate 90% and points only +2.

7 Likes
54

It’s not territory estimation, it’s more like prediction of average game result. If someone has a weak group it’s gonna be counted in this prediction, if someone doesn’t place their opening stones the best way, it’s counted in. If someone has a loose moyo that’s not territory yet it’s gonna be counted in. It’s not territory estimation, it’s better.

But we all knew that already.

I’m baffled at you saying it’s not reliable though. I’m pretty sure it’s taken as reliable even by strong players.

4 Likes
55

Back on the original topic of early side stones, one of the key ideas in go is that you cannot move your stones once played, so you want to put them in places which you later welcome rather than regret in terms of their efficiency with stones played in the future. As the corners are the anchors of the go board with their distance from 2 edges, you will likely end up playing in the corners (approach opponent’s corners if not starting there yourself) later, and then these corner stones will want to make extensions on the side. But if you have already played on the side then you can’t choose where to make an efficient extension based on where those corner stones are, you already made that decision in moves 1-8. So either you end up with an inefficient side stone, or you limit your choices of sequences you can play in the corners to ensure the side stone ends up efficiently placed, which severely limits your joseki choices and a skilled opponent could exploit this.

5 Likes
56

Yeah, and funny thing is that the deep endgame is the part where it’s the worst relative to human abilities. Not because it’s objectively worse, just relative to humans. Humans can be good at counting accurately in the deep endgame and therefore be competitive with or even better than bots then. But early in the game humans are terrible compared to bots at estimating early advantages and disadvantages, so there bots completely dominate (except in the presence of specific blind spots, etc).

Part of this is because humans switch methods as the game goes on. In the opening, all you can do is holistic intuitive feel, and that’s not very accurate. But late in the game as borders near completion, a strong player can switch methods from just intuitively feeling a position, to instead consciously enumerating points and quantifying everything, and gain a lot of accuracy.

Bots only ever use their intuitive feeling, or something loosely analogous to that, even once the position becomes precisely countable. Hence you can get funny results like deep in the endgame when Black is destined to win by 1.5 points and any strong player spending the time could solve the endgame and absolutely prove B+1.5, maybe the bot estimation is some funny fractional value like B+1.9. This is because unlike humans, the bot is not counting the position. The bot is doing the same thing it does in the opening - glancing at the whole board, and declaring based on intuitive whole-board feeling that it thinks the degree of advantage in points is something around B+1.5 or B+2.5, without ever counting it. This intuitive feeling is just so absurdly good that it is usually within fractions of a point of the correct answer you would obtain by explicit counting.

This is also why for large score differences it can be off by a whole point or more. E.g. as far as at-a-glance whole-board feeling goes, “black wins by 70.5” feels very similar to “black wins by 71.5”, they’re both “black wins by a lot”. It’s still usually not far off despite that.

So in the opening and early parts of the game, when explicit counting isn’t possible, such a good intuitive feeling just totally dominates what humans are capable of. We’re just left with bot intuition as being basically the most accurate thing by far we have at the moment.

(Again, caveat being that the position is “normal” and free from extremely sharp imminent tactics or blind spots. Remember, the entire thing the deep learning revolution has done is turned the comparative advantage of bots vs humans on its head! Strength of the machine is fuzzy holistic feeling and high level intuitive judgment, strength of humans is sharp tactical calculation and concrete mechanical reasoning).

8 Likes
57

And yes, KataGo is trained to explicitly estimate score, it is not simply a transformation of winrate.

As others have stated earlier in this thread, a reported lead of +6 roughly means “6 points is the amount of free points you would need to give the opponent to make the bot judge the game to be even, or to be uncertain and not prefer one side over the other”.

10 Likes
58

I think at least some pros do actually try a more counting approach even early in the game.
In the book “Nie Weiping on Go” (after a load of life story and communist propaganda) he counts an opening, ~move 50, position with enumeration of likely solid territory plus some hand-wavy quantifying of other moves values like a lone 4-4. I tried such an approach a bit myself with some of my OGS games, but was never convinced the error bars on my answers were smaller than the difference between black and white I got. But Nie was one of the best in the world at this stage of the game so maybe his results had justified accuracy. I’ve not revisited it since AI, could be interesting.

4 Likes
59

Yeah, I originally had a draft of that post that tried to address that too but I figured that was getting into too much detail so I went with something simpler. I agree, for a counting-based approach earlier in the game, it might work for someone better than alternatives depending on how their brain works and the way they conceptualize a position, but also yeah, I think you still have the challenge of accurate judgment in all the components that you’re adding together. The silliness of bots is being great at the judgment part… and then somehow kind of only passable at the “actually counting and/or numerically adding it up” part.

5 Likes
60

The game with @shinuito has ended. There are quite a few comments in the game while playing, as well as some variations:

All in all a very fun game and I think that it created a playable board.

3 Likes
61

On a very different note, since we are here, this is how the DGS test game is going. Unlike the game with @shinuito and the games with @Lord_o_o_Spoon in this case the opponent doesn’t know it is a test and this is an official ladder game:

Screenshot_18
Screenshot_18584×584 107 KB

I will probably lose the game (I’ve barely won against dan players only a couple of times), but so far I am very happy from the opening. Edit: The opponent tried @gennan 's approach of disrupting the full completion of the sideseki.

6 Likes
62

The fun continues. :slight_smile: My opponent this time is very keen on fun/unusual oppenings and not playing joseki, so here is how the first moved turned out. I think that this time this is in my favour since the side stones are already countering the usually influence-based opening of White.

6 Likes
63

Honestly, this whole experiment has been utterly fascinating to me. I’ve really enjoyed the challenge of playing against this, and indeed, its prompted me to be a tad more flexible in my approach to Go.

7 Likes
64

Sideseki experiment with White this time. It went pretty well:

A very interesting game overall with some very nice fights :slight_smile:

Personally I am starting to like this a lot since in a lot of games my main problem is that I urge the opponent to make too much influence and then I struggle to reduce it. But with the sideseki, there are already stones there for my convenience.

4 Likes
65

I uploaded the games on GoKibitz for anyone that likes that format and have a more open-style review:

1 Like
66

I seem to have answered the wrong post, I meant to reply to someone who still thought the score was derived from the win-rate

If I remember aright, KataGo can list an estimate of who is likely to end up owning each spot on the board. Something like this presumably underlies AI Sensei’s ability to estimate not just how many points a move losses, but also how much it loses at each spot on the board.

It shows these losses as an overlay of red, deepening as the loss increases, which can be very helpful in gaining a quick impression of why a move is sub-optimal.

This is why I think “converting % to score” is mistaken.