AI adjudication

I don’t know how big a problem disagreements about scoring or refusal to accept a score as a trolling/cheating tactic are, but I wanted to float an idea about how to resolve these types of deadlocks without taking any moderator’s time. (And maybe also deal with timed out games, which is probably a much more common annoyance.)

In Backgammon, software has practically solved the game, and so it’s common on their online servers to use software to adjudicate games where one of the players times out due to a disconnection. What I’m wondering is whether Leela Zero has gotten strong enough that we can use it to adjudicate games, at least between kyu level players. (And if there are in fact big problems with scoring disagreement or stalling, I suspect they exist mostly at the kyu levels.)

My idea is that when a certain level of disagreement or stalling is reached (say, when a game goes into the scoring phase for a third time) either player can shut the game down immediately and send it to Leela Zero to decide who won the game.

This might also be useful for deciding how to handle timeouts for rating purposes. The cases where the player who timed out was clearly losing vs. clearing winning vs. within some margin of equality could each be handled differently, with the goal of reducing any ratings distortion that might otherwise occur.

Anyway, maybe these issues are not big enough to warrant the effort and expense of getting Leela Zero running in the cloud and connected with the main server, but I thought it would be worth suggesting for future consideration if the problems ever grow or as the effort and expense go down. (Also, I for one want to go on the online record as often as possible as welcoming our future artificially intelligent overlords.)



  1. Get advantage
  2. Time out
  3. Profit
1 Like

I never suggested that’s how it might work. What I actually said was that Leela Zero would allow OGS to handle timeouts differently depending on the game state. So, for example, in the scenario you laid out, the game might simply be annulled, while if the timed out player was losing by a lot, it might be considered a loss. But I didn’t specify any particular outcomes because I felt the more important point was that in general, a Leela Zero judgment would give you more information that you could use to further whatever goal the site might have about how to handle timeouts. More information is always better than less, right?

That could be a good idea if the AI is reliable. But nowadays strong AI like fine art is still not 100% right on semial right? I think if refusing to pass becomes a very common issue(which I don’t think it is that bad) then it might be a good idea to use it as last resort.

To solve the issues smurph stated above, maybe disable AI until one of the player passed 10 times cumulatively, or when they played 300+ moves on a 19x19 board etc.

1 Like

This idea may have very limited use at the kyu level to resolve honest disagreements between players of goodwill. However, resolving the status of the game is not the most important thing when dealing with a troll or cheat. The most important thing in such situations is to identify the troll/cheat as such. This requires a moderator, who can issue a warning or a ban.

As for timeouts, it is unclear whether you are referring to correspondence timeouts (extensively discussed in another thread) or all timeouts. Timeout in a live game is part of the game and needs no adjudication. Disconnection is already well handled, I think, by the thunderbolt. Most thunderbolt losses are escapers (I base that on my own extensive observation of games). AI should not be adjudicating timeouts, because humans are not AI. One player may indeed have a won position based on AI analysis, but that doesn’t mean the player could really achieve the win. People make mistakes.

1 Like

Imagine a handicap game that times out after 50 moves. Black might still have a comfortable lead, that doesn’t prove anything about the outcome.

Even for even games, this system would favor players who are good in the opening over those who are strong in the yose.

1 Like

This topic was automatically closed 91 days after the last reply. New replies are no longer allowed.