Ytreza wrote: : i) there is no proof at the moment that black has the advantage. It might feel like it, but even between the strongest players, when looking at the archive, the balance is good. The AI playing white also manages to win 40% of the time even though it doesn't know forts yet.
ii) a win condition that requires to count the number of moves is utterly inelegant and impractical.
iii) if AI training stales, it's the problem of the AI and its developpers, not of the game. Ad-hoc rules like the fort are not easy to discover randomly by reinforcement. But stalling should not happen if draws (after 100 moves without capture) were penalised properly in training. The developpers might need to put more thoughts into their AI instead of applying a generic algorithm. Anyway, that shouldn't be the problem of the players.
iv) I have proposed in the big forum a way to speed up endgames by basically allowing black to claim the win when white is stuck.
v) it's now clear with the AI that nobody on this website, even the "GMs", are very good at the game. I struggle much more playing black against the AI that playing against e.g. Alex or Draganov.
Hi, Ytreza!
Unfortunately, I don't have access to the small forum, so I'm replying here.
1) Agreed — this is not proven at the moment.
2) Similar rules exist in chess, so I don't think such a rule would necessarily be a problem if the community agreed on it. In any case, I'm currently sticking to the existing Copenhagen rules and don't plan to change them on my own. When Alex suggested additional rule changes, I told him those ideas should first be discussed with the community.
3) Definitely. During training we can experiment with different reward shaping or anti-stalling penalties without changing the actual game rules for real players. Alex believes these issues should influence the official rules themselves, but that is his personal opinion, and I don't necessarily share it.
4) Yes, we will experiment with that. Idea to apply small penalties to both sides for extremely long games without captures during training only should be good I think. We also plan to try a smaller network on the Copenhagen variant, since it performed surprisingly well on the Historical variant.
