Low evaluations for clearly won endgames?
Mr-Twave opened this issue · 1 comments
https://lichess.org/tHNW0SD9#161
This endgame position comes from an lc0 vs Stockfish NNUE game.
Does the evaluation presented here returned from Stockfish make sense? The position is a KQvKR endgame which is clearly won, yet the evaluation returned by lichess is only +1, and I'm wondering whether this is a low-depth Stockfish issue, or another related issue.
One reason why I post this here is that any A.I. that train off lichess games are going to have a tough time training off these games if won endgames are 'drawish.'
Another point that I bring this up is that a serverside DTM tablebase query for 5-man positions would increase accuracy and save more compute time than using clients + Stockfish's time to evaluate endgame positions like these.
Thanks for reporting. Using tablebases does make sense, but was considered overall not worth it a while ago (lichess-org/lila#2290).