March 17th, 2019

LeelaChess прогрессирует

Между прочим, "народная" Альфа-Зеро, называемая ЛилаЧесс, доросла до уровня топ-движков, и уже играет с ними на равных и выигрывает (а до этого заняла второе место в чемпионате мира среди движков после СтокФиша):
losing 49.5 to 50.5
Stockfish may have won this superfinal, but it seems likely that Leela (NN based engines in general) may be the future of computer chess. There were many games where Leela played the opening much better than Stockfish, getting to positions where Stockfish realized it was worse or losing. Leela's closing skills are still lacking, it allowed Stockfish to escape many times. There was even one game where Leela failed to convert a 7-man tablebase win, despite having access to 6-man tablebases. Some of Leela's losses were a result of not looking far enough ahead and making a losing move, sometimes realizing it was losing immediately and sometimes many moves later. To summarize, it seems Leela is already better than Stockfish in some aspects of the game, and it needs to improve its skill in converting wins and calculating ahead. There are suggestions of a hybrid approach, who knows how much better engines can get.

Leela beat Stockfish in a 100 rapid game match with no book, the end result was 16-4 with 80 draws. Leela learned a lot about openings in its training, and it appears to be better than Stockfish from the initial position, at least in shorter time controls. (вот эта ремарка проясняет почему старший товарищ Лилы, Альфа-Зеро, имел доминацию над СтокФиш. Потому что они играли там без дебютной книги, а нейронки получается умеют это делать точно лучше.)

Эта, да-да, внутре у ней нейронка!