I’ve been successlessly bugging @anoek about this for a year… 
https://forums.online-go.com/t/please-fix-default-nz-rules-komi-for-ranked-9x9-and-13x13/
3 Likes
IIRC anoek is open to fixing small board komi, but would like to pencil in enough time to research and fix small board handicap at the same time. Maybe don’t hold your breath 
these things more often than not tend to be years rather than months (unless someone drops a researched, and data supported solution in his lap… that tends to give things a nice bump)
3 Likes
The data supported solution is called following the rules of the ruleset!
5 Likes
Most official rulesets do not specify komi for smaller board sizes, and certainly don’t stipulate a fair handicap system for them.
4 Likes
Even the number of handicap stones may be different on 9x9. When more advanced players accept my challenge, then see two automatic handicap stones, they get intimidated and leave. Two stones is a HUGE handicap on 9x9, just as standard komi is too large on 9x9.
1 Like
I think the point of this thread is to increase komi, because current 5.5 komi is way too low for most rulesets. It sounds counter-intuitive, but consider how massive handicap stones are on 9x9 - that’s how much greater the value of the first move is, as well.
But yeah, to address both handicap and komi at the same time seems like quite the undertaking to me, though I imagine there’s a fair amount of data from the games on this server at this point?
7 Likes
Does this support changing them from normal values?
Using the same non-integer komi for territory and area scoring is likely incorrect in any case.
Handicap stones are harder to play against on small boards, but this doesn’t mean they worth more than the usual 13-14 pointwise, nor that komi has different value. What differs is the value of being stronger.
Since the board is small with less moves for mistakes to accumulate, an extra N point is harder to win back. One rank strength difference amounts to 13-14 pts (= one stone) on 19x19, much less on 13x13, further less on 9x9. Otoh, the value of the first move (komi) is still half stone.
1 Like
At one time after komi was introduced (there was not always komi), all rulesets and board sizes used 5.5 komi
Eventually they decided that there was still an advantage for black so they lifted komi by 1 step. Territory could go up to 6.5 but area needed to go up to 7.5. This was agreed best for 19x19 but no decisions were made for the less serious sizes.
Now, since we are ranking such games, a decision is needed, even though there is no historical precedent.
I sympathize with the fact that changing the komi values would require rethinking the internal handicap system. I think this is the only valid reason to slow down the implementation of this fix.
In terms of the actual komi size, there is no evidence that komi should be smaller for boards between 9x9 and 17x17, even if the matter is not 100% settled. Even if rulesets don’t specify komi for smaller boards it shouldn’t be controversial to use whatever is the default komi is for 19x19 on such rulesets. The burden of proof with data should be on whomever wants to force a singular value of komi across all rulesets – not to those arguing that default values in 19x19 should be used for smaller boards. The whole point of having different rulesets is to be able to agree to disagree – if one thinks komi should be 5.5 then one should be able to play with the ING rules in a ranked game.
Yes i think this was not always the case but now thanks to katago i believe anoek is on board with komi going up on the smaller boards but would like to fix handicap at the same time.
TMK he is not committed to keeping 5.5 komi he just doesn’t want to upset the system multiple times if he can hopefully get it all right the first time.
3 Likes
Well, one nice thing would be, for 9x9, a handicap half-way
between this site’s handicap=1 and this site’s handicap=2 .
Specifically, it would be random choice (just like
random choice of color for even games) between
(a) 2 stones, but the stronger player gets significant komi
and
(b) no stones, but the weaker player gets Black and small reverse komi
.
(When I play these via custom challenges, I do 6.5 for (a) and 2.5 for (b),
though Katago might’ve subsequently changed its view on what values put
those halfway between this site’s handicap=1 and this site’s handicap=2 .)
Even without statistical/historical data, we can guess from theory that 0 komi wasn’t correct (since moving first is an advantage). Similarly, 5.5 for both scoring systems is unlikely correct, since B has 1 point extra advantage in every second game under area scoring.
1 Like
So 5.5 is very different from 7 (where 7 is 7 while 5.5 means komi 5 with W winning ties
You are 100% right here. I mean 5.5 is the same as 6.5 provided no odd-point seki in an area scored game. Scoring seki games fairly is still important though.
Isn’t the only reason we know that moving first is an advantage precisely because of statistical and historical data? If pros weren’t winning as black more often than white with no komi, we’d hardly be sure.
3 Likes
I pointed out years ago, in another komi thread, that the idea that komi should be the same for everyone is absurd on its face. The first move is practically no advantage to a raw beginner. This is why using statistics from pros, AI, and even high dans as validation for universal application is flawed.
2 Likes
Komi should likely be lower for double digit kyus, yes. But actually lower by less than you might expect.
Moves in the opening on 19x19 are worth about 14 points per stone area scoring, this is naturally why the practical fair komi is about 7, because it compensates black’s on average being half a move ahead of white. Under territory scoring, moves are 13 points on average since it doesn’t claim the point of territory under the stone itself (indeed, the “miai counting value” of endgame moves is precisely 1 point less under territory scoring than area scoring), this corresponds to practical fair komi being about 6.5.
There is also an different argument based on the parity of final stones that leads one to conclude why fair territory scoring komi should be 0.5 points less than fair area scoring komi - the fact that these two arguments give the same conclusion is not a coincidence, they both derive from the same underlying mathematical principles.
Anyways, in general, if opening moves are worth N, one would expect fair komi to likely be around N/2.
From first principles, the reason to expect fair komi to be lower for worse players is that for them, N is lower. They play worse, so they extract fewer points per move, so N is effectively lower, so fair komi ~= N/2 is lower.
But the amount by which players play worse is easy to overstate. Even kyu players extract the vast, vast majority of value on almost every stone they play. If you give up on average 2 points per move over the first 150 moves, that’s 150 points (you play half the moves), which is already a rank gap of like 10-12 ranks. Each doing 1 point worse per move corresponds fair komi being another 1/2 point lower.
So on this basis, we should be a little surprised if the ideal fair komi averaged across 1 kyu players turned out to be more than 1 point lower than for pros, and if the ideal fair komi averaged across 10 kyu players turned out to be more than 2 points lower than for pros.
The next-level of consideration is to consider when mistakes are made and the nature of mistakes, because mistakes are not equally distributed through the game. Mostly people don’t make huge mistakes in the opening, the biggest mistakes only start from midgame onward.
Imagine modeling the game as both players taking turns grabbing coupons that are each worth a different fixed number of points at scoring (and maybe sometimes taking those coupons reveal other coupons, or change the values of other coupons, etc). If it turned out that worse players grabbed equally-high value coupons in the opening as pro players, they were just worse about identifying and grabbing valuable coupons later, then the average fair komi would actually be the same! N in the opening would be the same since that’s not where the mistakes are.
That’s not the case in Go though, you can retroactively make your earlier coupons (i.e. stones) worse by playing in a way that doesn’t leverage their full value - i.e. even if you played objectively the right move, if you are going to reliably under-leverage a given earlier stone, then the earlier stone was effectively lower value. Probably some amount of point loss is attributable to this. But also probably plenty of the point loss is more like in the coupon game, losses idiosyncratic to midgame or endgame blunders and not actually decreasing the average value extracted from the opening move. (I.e. suppose we had various pairs of positions A and B where each A was objectively 13 points better than each B due to an extra stone, and we had a random population of kyu players play out each game. Across many such pairs of A and B probably the average difference between A and B would end up less than 13 points due to the kyu players not making use of the extra stone optimally. But probably also some of the mistakes made would be idiosyncratic blunders equally likely in both of A and B, and those blunders would not contribute to decrease the difference between A and B below 13 points).
So actually we might expect the 1 kyu fair komi to be somewhere well between 0 and 1 point lower than for pros rather than a full 1 point, and the 10 kyu fair komi to be somewhere well between 0 and 2 point lower than for pros rather than a full 2 points.
None of this theory is a replacement for data of course! It’s only some theoretical guidance for the magnitude of the effect that we might reasonably expect to look for or to find in data. In particular, it would not be a surprise in the slightest if from SDK and stronger, practical fair komi for humans turned out to always be within 1 point of that of pros.
Edit: fixing some really bad math typos.
3 Likes
Well, somehow the strategy of B passing first move (thereby turning W) never really took off, even before komi.
But even if go were invented today with 0 statistics collected, still there are theoretical reasons for guessing correct komi. Hexahedron beat me to the coupon argument, but even without that it is no surprise that having an extra stone on the board in every other turn/position cannot be too bad. And if moving first wouldn’t be an advantage, then (assuming no negative moves in the opening) a first move at A1 wouldn’t be inferior to a star point (since moving first equals to moving second after a 0 point pass).
I doubt that before-komi players were unaware of this (see historical handicaps of rotating B) - they just didn’t attempt to “fix” it directly, maybe for reasons of aesthetics, tradition or respect for the pure game.
I don’t think I’m understanding you correctly here.
My understanding for why Komi should be less for weaker players, according to this argument, is basically that weaker players make more bad/random moves, that end up devaluing their previous moves.
I’m following this far, but where you lose me is where I’m inferring that Black’s later moves devalue their previous moves more than White’s later moves do their earlier ones.
Wouldn’t it be reasonable to assume that both players, on average, would make the same number of devaluing moves?
Out of my 2000 games on Goquest at about 1700 elo (or whatever they use; I’m 3.5-ish kyu on OGS), my win-loss ratio is close to even (1037-956), and 2000 is a rather small sample size, I’d think. I’d be surprised if this ratio didn’t trend even across most or all ranks.
Edit: wait I think I get it all of a sudden. The issue is that Black can devalue their first move, but it’s impossible for White to devalue their Komi? Perhaps this really only applies on larger board, though? 7 Komi area scoring on 9x9 seems appropriate to me, but then that’s the gut feeling of a kyu player 
1 Like
Exactly, black has a bit more of their net worth in the form of “future value of stones on the board” whereas white has more of their net worth in the form of komi. So if both players have some tendency to devalue their earlier stones, probably on average black suffers from it a bit more.
But yes, under reasonable guesses, komi should be almost the same for everyone except perhaps DDK players.
In the very far limit of random play (random play is astronomically worse than even beginners), I once did an experiment that found that fair komi was about 1. In particular, 19x19, area scoring and random play but no self-eye-filling or wasting by playing on pass-alive groups, and passing only when out of other moves, the fairest komi was 1. Loosely suggesting a move value of 2 points for the opening move - maybe Black’s random play gained about 1 extra points of value on top of the pure area scoring value of 1 point that the stone itself had (presumably the value coming from that stone randomly connecting some groups or filling white’s liberties in some of the games).
EDIT: Looking up a bit more - by mean score, white was only 0.7 points worse on average on the board, rather than a full point, 1 point komi was just the fairest discrete value. So that would correspond to black doing about (0.7 - 0.5) * 2 = 0.4 points better for the extra move than its pure area scoring value - much less than a full point, but still not nothing.
4 Likes
I think firstly the coupon argument, at least anything involving 7.5 → 15 points, 7->14, and 6.5->13 points already assumes we’re firstly playing a game with komi, and secondly that the game isn’t fair without komi. A new player to the game of Go, and seeing the size of the 19x19 board might expect that maybe there’s a first turn advantage, but certainly given the size of the board, there might not be one. Maybe with 0 komi the theoretical best result is a draw, even though that would require perfect play to have eye vs eye seki or something else in area scoring.
If the first move was worth N=0 points then N/2 is still zero
Secondly it probably is the case that playing at A1 is worse than passing At least some initial estimates of Katago seem to suggest that, but maybe super high playouts Katago doesn’t care. However since the 1-2 point ends up being quite important in many corner life and death puzzles, it wouldn’t surprise me if A1 is worse than passing
My guess is that people don’t pass on the first turn because they actually want to play the game Certainly one way to “win” is not to play, but that’s not much fun.
Anyway, my point is that you predicate your point on the assumption that Go is unbalanced and needs komi for the second player, but most of the intuition from this either comes from statistics (pro games or the millions of training games ai does to learn to evaluate positions) or historical data (like pro games etc). I think even Berlekamp’s idea of coupon go was just to try to help quantify how pros interpret the value of moves in Go in positions other than the endgame.
Edit: Another thought I had but forgot, was that if current top Ai can give 2-3 stones to professional players then it could easily have been the case that in fact Black is the one that needs reverse komi 
That doesn’t seem to be the case though from the various AI evaluation functions though from their self play.