I couldn’t for the life of me, but I was wondering if strong players can do that.
Other than the actual scoring phase, are there any telltale signs, what a pro is going for or avoiding, that can hint if it’s area scoring or territory scoring for example?
No. The gameplay before dame is to all intents and purposes identical.
If a ref pauses the game because stones are on the table, it might be Korean rules
I may be wrong, but isn’t a seki scored differently under different rules? So, for example, if a ko threat could lead to a seki, could it matter in any way which ruleset it is?
The life and death status of a group can change according to whether a ruleset allows superko or not.
So, a pro would do what differently if superko is allowed? Try to kill that group or not?
Most sekis have an equal number of eye points for either side, so scoring as 1 for black and 1 for white gives the same as 0 for black and 0 for white. Occasionally you do get asymmetric number of eye points in sekis, in which case yes the value of making the seki can differ, but in such a situation it’s probably a large seki worth like 30+ points in a fight, so it’s not like there’s another 29 and 31 point gote endgame move sitting there that you could take instead, and would make a different choice in territory vs area scoring because the seki is 1 point difference in value.
On 9x9 you might be able to tell if you recognize a move that maintains the draw in area but loses in territory
Probably when it’s at the point where you get signs it could be a certain ruleset, it’s probably already confusing pro players what’s going on in the game
like
or
I’m trying to remember a specific game where the players also didn’t remember how korean rules worked and whether a group was dead or not. It might come back to me
Here are some fun cases on small boards where seki scoring does end up mattering:
7x7: KataGo Opening Books - Deep Seki Rules Differences
9x9: KataGo Opening Books - 9x9 Highlights and Discoveries (see diagram 30 and 31)
Since some joseki end with seki, it would not be inconceivable to have a joseki that leads to a fair outcome under one ruleset but that favor one of the players (by 1 or 2 points) under another ruleset - but I never heard of any such examples! Since seki is rare in the first place, and unbalanced seki even rarer, maybe it just doesn’t exist.
Can anyone think of a joseki that ends in a seki that is scored differently depending on the rules?
I suppose one thing that might hint it is if in endgame one player is playing safe, and they would be winning by 0.5 in one ruleset but losing in another, that suggests they are in the former ruleset.
Isn’t there a ruleset that allows suicide moves? But I don’t know if it’s obscure or used in pro games.
yeah, NZD does, but it doesn’t come into play very often; even when it provides a ko threat, it’s often a bad ko threat
I don’t think pros are strong enough to account for a difference in superko rules midgame with live time control.
Besides, OTB pro games don’t actually use superko even with Chinese rules, so if a pro game would hinge on superko, it would turn out as no result in practice.
Yes there are such rulesets (like Ing rules and NZ rules), but very rarely such a difference would matter. And even if it matters, it would likely only change the number of ko threat threats by 1. I don’t know about any pro game where this affected the game result
In some very rare cases, possibly (if the players also grasp the nuance of some potentially complicated rules beasts that they find themselves dealing with), but almost always, the fundamental strategy of Go and the optimal choice for each move remains the same.
Things like joseki, fuseki, and (typical) life and death patterns are basically universal. I’m not aware of there being any example of Chinese rules joseki being different than Japanese rules joseki.
By the way, @elsantodel90 made the below post that lists the possible (but very rare) situations where the optimal move choice would vary between the common rulesets:
The last case listed in that post is the type of situation that I discuss in this thread:
I don’t have an example of that, but I think I do have an example where the rules affect the evaluation of an opening joseki in an amount that’s meaningful for strong bots, due to how an odd-dame seki affects the parity.
On an empty board, Chinese rules with 7.5 komi favors white much more than Japanese with 6.5. The reason is that in the absence of stuff like sekis or post-dame kos, due to parity we have (OTB = on the board, before komi)
(…etc)
JP B+5 OTB <=> CH B+5 OTB (case A)
JP B+6 OTB <=> CH B+7 OTB (case B)
JP B+7 OTB <=> CH B+7 OTB (case C)
JP B+8 OTB <=> CH B+9 OTB (case D)
(etc…)
And that means in Japanese 6.5 komi, black wins cases C and D and further, but in Chinese rules 7.5 komi, black only wins case D and further.
Suppose there is an odd dame seki. Now the board parity becomes even instead of odd, so we get:
(…etc)
JP B+5 OTB <=> CH B+6 OTB (case A)
JP B+6 OTB <=> CH B+6 OTB (case B)
JP B+7 OTB <=> CH B+8 OTB (case C)
JP B+8 OTB <=> CH B+8 OTB (case D)
(etc…)
Now suddenly in both of Japanese 6.5 and Chinese 7.5, black wins cases C and D and further.
This means that if there is an opening sequence leading to an odd-dame seki in Japanese 6.5 rules, then under Chinese 7.5 rules that sequence should typically favor black a little more, due to causing case C to flip.
Here is an opening joseki that sees real play that has an odd dame seki.
(Happy flying daggery messiness!)
We can try to validate by looking at bot winrates, and bot winrates I think reflect that this joseki really is different under the two rules due to parity!
In a plain opening with all 4-4 points occupied alternately by the players, KataGo doesn’t think this joseki is exactly even under Japanese 6.5 rules, but it’s close. For a recent b28 net, it looks like:
JP6.5 all 4-4 points blackwinrate 48% → Post-joseki blackwinrate 55% (+7%)
But how about under Chinese rules 7.5?
CH7.5 all 4-4 points blackwinrate 37% → Post-joseki blackwinrate 57% (+20%)
Bigger upswing for black, as expected.
Let’s explore different komi:
CH5.5 all 4-4 points blackwinrate 58% → Post-joseki blackwinrate 76% (+18%)
CH6.5 all 4-4 points blackwinrate 59% → Post-joseki blackwinrate 68% (+9%)
CH7.5 all 4-4 points blackwinrate 37% → Post-joseki blackwinrate 57% (+20%)
CH8.5 all 4-4 points blackwinrate 38% → Post-joseki blackwinrate 51% (+13%)
CH9.5 all 4-4 points blackwinrate 21% → Post-joseki blackwinrate 43% (+22%)
In every case the neural net still does think the joseki is good for black, but the increase is markedly larger on favorable-parity komi than on unfavorable-parity komi, again as we would expect.
Note also though that the above winrates indicate that KataGo’s sensitivity to parity is not perfect. The all-4-4 points position shows very clearly the normal effect of parity as you go from CH 5.5 → 9.5 komi, but adding the odd dame seki does not invert the parity staircase but rather just makes it kind of smooth. And I’m a bit suspicious of whether the CH winrate deltas really are mathematically consistent with the JP winrate delta here, although I’d need to think about other possible higher-order effects on the score distribution.
But anyways, I think winrates being rough here isn’t surprising - games with this pattern or other odd dame seki are still only a tiny tiny fraction of all games. And to merely to estimate a binary random variable to within a few percent accuracy takes thousands of games, so with a tiny fraction, we’re looking at possibly hundreds of thousands of games needed statistically. And the neural net has to actually realize that this parity thing suddenly inverts in this uncommon case just by sheer weight of data (because such a net cannot reason abstractly and mathematically about it like we can and did above).
And of course, given how violent bot self-play games often are and how willing they are to engage in mass trades, I would guess often the odd dame seki can collapse anyways due to a sacrifice or ko and you return to the original parity anyways, so the bot kind of generally learns that odd dame seki means a bit of a blur of the two parities and that might actually be somewhat true.
Still, the overall takeaway I think still stands. If you’re at the level where the parity of the board and komi matters a lot (and bots are starting to get there, although clearly noisy and not perfect), merely an odd dame seki can cause JP vs CH rules to diverge in evaluation, and we do have a real joseki that sees play that is affected by this!
That’s some interesting data.
Some hypotheses that I would draw from those numbers:
To test those hypotheses further, I’d be interested to see some more data about the winrates of all-4-4 versus post-joseki with other half-integer komi sizes under Japanese rules, and also about this data with some integer komi sizes under Japanese and Chinese rules.
[*]
This corresponds to about 65 Elo difference, so the rank width at that AI’s level is roughly 850 Elo. You can compare that to a rank width of about 50 Elo around SDK level, about 100 Elo around 1d EGF level and about 300 Elo at pro level.
Or expressed as winrate percentage point per board point: ~0.5 pp/bp at SDK level, ~1 pp/bp at 1d EGF level, ~3 pp/bp at pro level and ~9 pp/bp at this AI’s level.
From those hypotheses and earlier posts about AI komi evaluation, I’d predict the following idealised results when translating black winrates to black point leads (black’s point gain from the joseki between parentheses):
JP3.0 all 4-4 points black +3.0 → Post-joseki black +4.0 (+1)
JP3.5 all 4-4 points black +2.5 → Post-joseki black +3.5 (+1)
JP4.0 all 4-4 points black +2.0 → Post-joseki black +3.0 (+1)
JP4.5 all 4-4 points black +1.5 → Post-joseki black +2.5 (+1)
JP5.0 all 4-4 points black +1.0 → Post-joseki black +2.0 (+1)
JP5.5 all 4-4 points black +0.5 → Post-joseki black +1.5 (+1)
JP6.0 all 4-4 points black =0.0 → Post-joseki black +1.0 (+1)
JP6.5 all 4-4 points black -0.5 → Post-joseki black +0.5 (+1)
JP7.0 all 4-4 points black -1.0 → Post-joseki black =0.0 (+1)
JP7.5 all 4-4 points black -1.5 → Post-joseki black -0.5 (+1)
JP8.0 all 4-4 points black -2.0 → Post-joseki black -1.0 (+1)
JP8.5 all 4-4 points black -2.5 → Post-joseki black -1.5 (+1)
JP9.0 all 4-4 points black -3.0 → Post-joseki black -2.0 (+1)
