Ancestors

Toot

Written by gmmds on 2025-01-27 at 20:50

[#]reengineeringtheclassics #siliconroad #lichess #chess Final game of the Leela Knight Odds - GM Joel Benjamin match with Joel leading 4-3! Join me in 10 minutes at 21:00 UTC tonight at https://m.youtube.com/@SiliconRoadChess for the thrilling conclusion!

=> More informations about this toot | More toots from gmmds@mastodonapp.uk

Descendants

Written by skiminki on 2025-01-27 at 23:03

@gmmds This was an excellent conclusion to the match!

=> More informations about this toot | More toots from skiminki@mastodonapp.uk

Written by gmmds on 2025-01-27 at 23:15

@skiminki Pretty cool! Somehow Leeks manages to make something of the most unpromising situations!

=> More informations about this toot | More toots from gmmds@mastodonapp.uk

Written by skiminki on 2025-01-27 at 23:21

@gmmds That resistance is something that seems to be very specific to Leela, probably due to the way Leela evaluates the positions using expected scores.

I remember that when I practised KBNvK mating, Leela was the only engine that created any serious resistance. In fact, used every trick in the book and then some.

Stockfish and other engines pretty much just played with great apathy, because all moves were losing anyway. (Or maybe SF NNUE is now better, haven’t tested this in a while.)

=> More informations about this toot | More toots from skiminki@mastodonapp.uk

Written by gmmds on 2025-01-27 at 23:24

@skiminki No it’s true. You can definitely make a draw more easily against Stockfish with good preparation - endings it considers drawn are played very loosely somehow. It presses for a while and then gives up! Leela like this is horrific to resist against even in the simplest endgames!

=> More informations about this toot | More toots from gmmds@mastodonapp.uk

Written by zaunkoenig on 2025-01-28 at 05:18

@gmmds @skiminki Well, that's obviously a "feature" of a tree search Stockfish is using. If there are draws everywhere, the scores for all moves are the same, because it finds the correct drawing path for every move and does not keep wrong paths. A Monte-Carlo approach (lc0) struggles to find the correct path first, but after enough games it starts to slowly converge to the "true" score. The more complex a position, the longer that takes. Wrong attempts are still in the score.

=> More informations about this toot | More toots from zaunkoenig@mastodon.online