We are experiencing a rise of questions relating to the end of the game and scoring phase, usually phrased as bug reports.
New (and even SDK) players are not aware that the end game needs to be played out.
Users do not recognize the difference between score estimation and final scoring.
The classic cheats mark live groups dead and are disagreeable.
Disaffected players prefer not to call for moderator help, but take to these forums after the fact.
Some have suggested solutions around the idea of “judgement by bot” in recent topical threads. I want to expand on that to offer a solution that is principally easy to operate and easy to understand.
Let’s have a new “OGS Ruleset”, default for ladders and site tournaments, that basically works like this:
Once both pass, the game is over. No resumption.
A strong bot scores the game.
The bot looks at the position from each side to move next.
If the “loser” has a winning move to turn the game around (i.e. result depends on who’s turn it is), both players lose the game.
This procedure works the same as manual scoring for the finished game, but less prone to errors or manipulation.
Even in unfinished games with open borders, the game finishes seamlessly according to users’ expectations.
Many weird scenarios that arise out of middle-game 2x-passes or people not fixing severe weak points are handled gracefully. Both players lose an unclear position, so there is incentive to finish the game properly and disincentive to try mischievous tricks.
Let’s preempt the rebuttal: that it would be preposterous to introduce bot scoring because it does not conform to the rules of Go, that two players must agree on the final position.
It has to be pointed out that the existing rules implementations are already not conformant, especially the “Japanese” rules on any Go server. They vitally neglect to implement hypothetical play. Let’s extend this creative liberty a few more steps and give it a distinguishing name.
I’m not against some creativity but i see some problems coming with a “automatised endgame.”
Even if the procedure requires a mutual agreement by both passing:
The potential winner is going to have a chance to deprive the other of opportunities of reversing the result. That’s not go for me as a go player skill include the consistency to keep an advantage to the very end of a game.
The potential loser may decide that because he don’t see any issue by himself, to give a chance to the AI (that you don’t see) to deny the result for his opponent instead of simply resigning himself. That’s like an encouragement for legal bad behavior.
The fact that rules are not well replicated is not really an argument to create new one as there is still ways to replicate them better and the japanese one maybe not the best choice to use as an argument. It’s not by chance that bots use mostly the chinese rules.
I think you made some good points - to a proposal different from mine in which “forced AI arbitration” is a thing.
If I pass, what grounds to complain do I have afterwards that I was “deprived of my opportunity to reverse the result”? If I had this opportunity, I deliberately gave it away.
If I am losing, should I try a pass, hypnotize my opponent to follow, then hope that actually contrary to my own understanding, I was actually way too far ahead to be caught up and will be awarded the win? Or should I just make a move, start a fight, make something happen…?
The argument for new rules is that the existing ones have fixable problems. Even if the implementation was accurate, that would not improve on the scoring confusions specifically.
(Pseudo-) japanese rules are widespread as a matter of fact. Not my fault.
Well that’s a problem because you don’t rely on yourself anymore but on an external help, the AI. Even if, as in your first proposal it results in a lose/lose situation in the case the luck is on your side, i prefer the honesty of a resign.
I like this example! It’s way more subtle than the obvious open cuts in the other thread.
But yes, if you leave a way for your opponent to seki your corner, even if your opponent is too weak to find it, you have failed to prove that you are able to clearly beat them. I would advise black to take it as a valuable learning experience.
This example is an extreme outlier, obviously, and one can make an argument that the board should be scored “as finished” if it “looks finished”.
Well that was a main argument in the topics earlier.
We can oppose a rational argument now (like for truth, or teaching goals, progress) but you know we have different public
I’ m not sure if we should care of the ruleset too as white has 1 point in the seki (an eye) in chinese rule. I was going to forget that black has one too.
There are no prisoners (if using the japanese rule) because same numbers of stones.
In aga rule black pass before white play for seki which will result in 1 black prisoner
The proposal involves two judgements, one with black to move and one with white to move.
If the same color wins in both judgements, that color is declared winner.
If black wins on black to move and white wins on white to move, even if just by 0.1 points, that would be the “both lose” situation.
This would be a close endgame situation where both forget to take the last decisive point.
white to move → w + 4.2
black to move → b + 0.1
will it be a loss for both? I’m asking because I know that AI evaluation is not always 100% conclusive (otherwise values like 0.1 would not be possible), so if the AI estimation is close to 0, its output value can very well be on the wrong side compared to the “theoretical optimal play result”.
However, modern bots are rarely more than 0.5 points off on a finished board. It would take some quite extreme position where a bot like KataGo gives white -0.6 for not playing dame.
I think by very late endgame, and the end whereby the ai wants to pass (no dame to fill etc) I imagine one could probably come up with a rule like “round to the nearest half odd integer” if using komi x+0.5 with x and integer.
Like W+0.1 and similarly small amounts might round to W+0.5 and B+1.2 would round to B+1.5 etc.
I think the numbers like W+0.1 are simply due to the fact that the ai is trying to fit a function, by the end of the game it’s really simply a scoring function that takes in a board position and counts the points, except that isn’t what the Ai was trained to do. Instead it needs to predict with some certainty the outcome, from most of the time unfinished positions. So the uncertainty and lack of perfect approximation leads to numbers (score estimates) which can’t be real on the board + komi scores.
In fact, humans also do this with unfinished positions in endgame, assigning fractional points to the size of moves (not just half integers, but 1/4, 1/8 etc) but more bizarrely infinitesimals which again need to be rounded in some way to figure out which player wins and by how much.
Anyway, I don’t think it’s unreasonable to suggest a simplified mode, or a new rule set which auto scores. I mean we more or less have that with bot games right? I guess one needs to pick a rule set to base it off like New Zealand, Chinese, Japanese etc (area vs territory). One maybe could use a modified scoring algorithm based on a bot like Katago, such as what Vsotvep was suggesting
It is an interesting idea to apply autoscoring to things like ladders and tournaments (the both lose in particular).
Whether people would be happy with both lose scenarios I’m not so sure. So whether it would be popular outside of games where it was forced (suggested ladder etc) I also don’t know.
It can handle any case gracefully in the sense that there is an easy to understand resolution to every user action, no matter how misinformed.
Nothing changes about the fact that the game needs to be finished for proper scoring.
My proposal offers the proper incentives for that.
Calling the game early still depends on your opponent to agree, and even if they do, it will probably result in “both lose” unless the position is overwhelming.
This adresses the problem cases from the first post.
No, white failed to gain something in the corner against black, while black failed to win this position against the AI.
Sure, we could introduce the rule “The winner is who can prove, that they can clearly beat their opponent”, but I assume the use of the word “clearly” would not help the goal to achieve a simple ruleset.
This example is the only example in this thread. Calling it an “extreme outlier, obviously” seems harsh.
And when both players pass, aren’t they saying, that the board “looks finished”?.
Especially when playing against bots it automatically scores with a bot that makes some inaccurate decisions before i got unfair wins and loses against bots so that would be better for bots
