Leela Chess Zero is the open-source community reimplementation of the approach DeepMind used for AlphaZero. Rather than searching as many positions as possible, it uses a deep neural network to evaluate a much smaller number of promising lines, guided by Monte Carlo tree search.
The network was trained by playing tens of millions of games against itself, starting from nothing but the rules. Nobody told it that bishops are worth roughly three pawns or that doubled pawns are usually bad — it worked out what mattered from the results of its own games.
The result plays noticeably differently from Stockfish. Leela is often described as more "human" or more positional: it will accept long-term structural compensation for material in ways that traditional engines evaluate as simply losing. Watching the two disagree about a position is the most interesting thing you can do with two engines.
What makes it distinctive
- ▸Neural network evaluation with Monte Carlo tree search, not alpha-beta
- ▸Searches far fewer positions per second than Stockfish and compensates with better judgement
- ▸Frequently disagrees with Stockfish about positional sacrifices