4 ms·

> These computer AI chess programs have weaknesses. > ... have collected statistics on which openings will lead to situations where Stockfish (or LeelaZero, or

> These computer AI chess programs have weaknesses. > ... have collected statistics on which openings will lead to situations where Stockfish (or LeelaZero, or Komodo, etc. etc) will inevitably fail. In case of Stockfish and other "conventional" engines, yes. However, with engines like LeelaZero those weaknesses gets much harder to exploit, since you can create your own version of Lc0 by training a new model with tensorflow.

dragontamer 7y ago
LeelaZero struggles with endgames. I'd probably use LeelaZero through the beginning and middle games, but switch to Stockfish whenever I feel like a tactical situation comes up, especially if the position starts to drop down to 10-man or 9-man. 7-man Syzygy Tablebase are less than 10TB, a RAID-array of M.2 drives will access those quickly. If Stockfish evaluates any 8-man trade, it will simplify into a 7-man situation and you have guaranteed perfection after that. No one knows how to integrate the endgame tablebases into LeelaZero by the way. So if we both play LeelaZero for the first 40 moves, but I switch to Stockfish + 7-man Tablebase for the ending, I'm probably going to beat you. LeelaZero's endgame really is quite weak. --------- If you are a programmer, and can figure out how to integrate the Neural-Net evaluation into MCTS to evaluate tablebases, you probably would do great. Its currently unsolved (to my knowledge). It seems solvable... I don't see any reason why MCTS couldn't have a if(popcnt(bitboard) < 8) doTablebase() statement somewhere. The question is how to do it as quickly as possible, and how to do it asynchronously and integrate the statistics to the rest of the MCTS tree. BTW: A research team has created the 8-man Tablebase, although it takes probably 100s of TB to store. This means that 8-man Tablebases will remain as Hard-drive only for now.

dragontamer 7y ago

LeelaZero struggles with endgames. I'd probably use LeelaZero through the beginning and middle games, but switch to Stockfish whenever I feel like a tactical situation comes up, especially if the position starts to drop down to 10-man or 9-man. 7-man Syzygy Tablebase are less than 10TB, a RAID-array of M.2 drives will access those quickly. If Stockfish evaluates any 8-man trade, it will simplify into a 7-man situation and you have guaranteed perfection after that. No one knows how to integrate the endgame tablebases into LeelaZero by the way. So if we both play LeelaZero for the first 40 moves, but I switch to Stockfish + 7-man Tablebase for the ending, I'm probably going to beat you. LeelaZero's endgame really is quite weak. --------- If you are a programmer, and can figure out how to integrate the Neural-Net evaluation into MCTS to evaluate tablebases, you probably would do great. Its currently unsolved (to my knowledge). It seems solvable... I don't see any reason why MCTS couldn't have a if(popcnt(bitboard) < 8) doTablebase() statement somewhere. The question is how to do it as quickly as possible, and how to do it asynchronously and integrate the statistics to the rest of the MCTS tree. BTW: A research team has created the 8-man Tablebase, although it takes probably 100s of TB to store. This means that 8-man Tablebases will remain as Hard-drive only for now.