📝Greg Strong wrote on Sun, Feb 16, 2020 12:22 AM UTC:
Thank you for the help. It is indeed pretty strange.
First, I should clarify a couple of things. The problem isn't specific to KRK - I know it appeared in other situations as well, but KRK was an easily reproduced and extreme example. Also, the special endgame code isn't specific to KRK, it is really KxK (King + plenty of material against bare king, probably the same idea as your bare king eval.) This KxK code comes from Stockfish but adapted to do the right thing on boards of any size. The other specific endgame evals I have so far are KRKP, KRKB, and KRKN (also from Stockfish.) There is also some special handling for KPK which overwrites the eval to return draw scores on certain positions that are known to be draws (this is custom code based on Wikipedia's KPK article - the Stockfish KPK code uses a bitbase which doesn't help me given variably sized boards.)
The null move was a good idea, as well as draw scores getting into the TT, but neither of these seem to be the cause. Null move is disabled with low enough material and disabling draw-by-repetition and/or 50-move rule didn't help.
With a fair amount of expirementation, I found another "solution". If I change the TT cutoff condition at PV nodes from hash depth >= depth remaining to hash depth > depth remaining that solves the problem entirely and my searches seem totally consistent. But I have "solution" in quotes because I'm not 100% sure this addresses the real problem either, but I think it probably does. Somehow something is getting off by one somewhere. I still need to understand it better because the underlying problem may be affecting non-PV nodes too resulting in weaker play that is not so obvious...
Thank you for the help. It is indeed pretty strange.
First, I should clarify a couple of things. The problem isn't specific to KRK - I know it appeared in other situations as well, but KRK was an easily reproduced and extreme example. Also, the special endgame code isn't specific to KRK, it is really KxK (King + plenty of material against bare king, probably the same idea as your bare king eval.) This KxK code comes from Stockfish but adapted to do the right thing on boards of any size. The other specific endgame evals I have so far are KRKP, KRKB, and KRKN (also from Stockfish.) There is also some special handling for KPK which overwrites the eval to return draw scores on certain positions that are known to be draws (this is custom code based on Wikipedia's KPK article - the Stockfish KPK code uses a bitbase which doesn't help me given variably sized boards.)
The null move was a good idea, as well as draw scores getting into the TT, but neither of these seem to be the cause. Null move is disabled with low enough material and disabling draw-by-repetition and/or 50-move rule didn't help.
With a fair amount of expirementation, I found another "solution". If I change the TT cutoff condition at PV nodes from hash depth >= depth remaining to hash depth > depth remaining that solves the problem entirely and my searches seem totally consistent. But I have "solution" in quotes because I'm not 100% sure this addresses the real problem either, but I think it probably does. Somehow something is getting off by one somewhere. I still need to understand it better because the underlying problem may be affecting non-PV nodes too resulting in weaker play that is not so obvious...