![]() It can draw completely winning positions because it thinks every move is winning, including some that are not. Lc0 is not trained well for winning endgames.This is 3 orders of magnitude difference, yet they evaluate to a similar depth in most positions (around 35), indicating that Stockfish sees a much wider tree, better evaluating more tactics positions, whereas Lc0 sees a deeper tree, evaluating deep endgames and positional weaknesses using its vastly superior heuristic. On the hardware in these tournaments, Stockfish evaluates around 60,000,000 nps, whereas Lc0 evaluates around 60,000.In head to head matches, typically Lc0 wins reliably, but Stockfish is better at beating weaker engines (avoiding draws more successfully), so Stockfish still wins many engine tournaments.However, they are very different engines, based off very different methods and strengths, and so the goal is to combine those strengths. If you need windows or Linux binaries, you can contact me at Hybrid Projectīoth the Lc0 chess engine and the Stockfish engine are extremely strong, and have been dramatically improving due to improved heuristics for Stockfish, and improved training for Lc0. ![]() ![]() CodeĬode is publicly available on Github. Lc0-stockfish-hybrid is a (mostly) UCI-compliant chess engine based off two separate projects, Stockfish the leading traditional chess engine, and LC0, or leela the leading open source neural network chess engine based off AlphaZero.
0 Comments
Leave a Reply. |