chess

A chess engine written in Rust, using the classical alpha-beta pruning algorithm for best-move selection.

Installation

Clone this repository, and then run:

cargo install --path .

Once installed, you can run the engine with chess, so long as you have the chess binary in your PATH (e.g. export PATH="$PATH:$HOME/.cargo/bin").

Usage

$ chess --help
chess 1.0.0
A classical chess engine implemented in Rust ♛

USAGE:
    chess <SUBCOMMAND>

FLAGS:
    -h, --help       Prints help information
    -V, --version    Prints version information

SUBCOMMANDS:
    count-positions            Count the number of possible positions for a given `--depth` (default: 4), and
                               reports the time it took to do so. By default, this searches all possible positions.
                               The routine can be run with alpha-beta pruning by selecting `--strategy alpha-beta`.
    determine-stockfish-elo    Determine the ELO rating of the engine at a given `--depth` (default: 4) and
                               `--starting-elo` (default: 1000). The engine will increment the Stockfish ELO until
                               it plateaus at a 50% win rate, at which point the rating is reported.
    help                       Prints this message or the help of the given subcommand(s)
    play                       Play a game against the computer, which will search for the best move using alpha-
                               beta pruning at the given `--depth` (default: 4). Your starting color will be
                               chosen at random unless you specify with `--color`.
    pvp                        Play a game against another human on this local machine.
    watch                      Watch the computer play against itself at the given `--depth` (default: 4).

Performance

Throughput

On an M1 MacBook Pro, the engine plateaus at around 45 million positions per second.

$ chess count-positions --depth 6
depth: 0, positions: 20, positions per second: 344827.5862068965
depth: 1, positions: 420, positions per second: 36910.09754811495
depth: 2, positions: 9322, positions per second: 924159.8096559928
depth: 3, positions: 206603, positions per second: 7812258.942751266
depth: 4, positions: 5072262, positions per second: 30028309.929195575
depth: 5, positions: 124135208, positions per second: 44738631.82584201
depth: 6, positions: 3320182902, positions per second: 44980862.13103737
total positions: 3449606737, total duration: 76.804796s, positions per second: 44913949.6054387

This is a pure depth-first search of all possible positions - no pruning is applied.

Alpha-beta pruning, which incorporates the engine's scoring heuristic to prune branches of the search tree, is used to search for the "best" move in actual gameplay. Using this approach, the engine plateaus at around 900,000 positions per second:

❯ chess count-positions --depth 6 --strategy alpha-beta
depth: 0, positions: 20, positions per second: 1125.4924029262802
depth: 1, positions: 420, positions per second: 32218.471923903035
depth: 2, positions: 4434, positions per second: 254359.79807251034
depth: 3, positions: 19494, positions per second: 588444.8200917653
depth: 4, positions: 501500, positions per second: 927077.8837852877
depth: 5, positions: 1122179, positions per second: 869289.6124067038
depth: 6, positions: 26710588, positions per second: 951893.6850460902
total positions: 28358635, total duration: 29.973702s, positions per second: 946117.1996705646

These figures vary by hardware. To achieve the best performance, make sure to use the release build, which leverages compiler optimizations:

Gameplay

To measure the engine's performance in actual gameplay, use the determine-stockfish-elo subcommand. This will increment the Stockfish ELO until it plateaus at a 50% win rate, at which point the rating is reported.

chess determine-stockfish-elo --depth 6 --starting-elo 2000

At alpha-beta search depth 6, you can observe the engine winning against Stockfish playing at a 2000 ELO.

Implementation details

There are numerous optimizations used to increase the engine's performance. This list is not exhaustive, but should give you a sense of the techniques used:

• The board state is represented as a 64 bit integer. This enables the engine to leverage the CPU's bitwise operations to quickly calculate moves, attacks, and other common board state changes.
• Alpha-beta pruning is used to quickly eliminate branches of the search tree that are unlikely to lead to a winning position. The move order is sorted in an attempt to prioritize the "best" moves first, so that worse moves come later in the search and are therefore pruned (and not searched entirely), reducing the search space/time.
• The Zobrist hashing tables are generated at compile time using the precompile build script. This hashing approach enables quick incremental hashing of the board state so that various computations can be cached (e.g. move generation) by the engine during gameplay.
• Macros are used throughout the codebase to improve the developer experience. See below for one example.

// The `chess_position!` macro is used to instantiate a board state from
// an ascii representation of the board. For example, here is the starting
// position:

let board = chess_position! {
    rnbqkbnr
    pppppppp
    ........
    ........
    ........
    ........
    PPPPPPPP
    RNBQKBNR
}

// This is used extensively in tests, where various positions are instantiated
// to exercise the engine's logic with.

```rust
#[test]
fn test_find_back_rank_mate_in_2_black() {
    let mut search_context = SearchContext::new(3);
    let mut move_generator = MoveGenerator::new();

    let mut board = chess_position! {
        ....r..k
        ....q...
        ........
        ........
        ........
        ........
        .....PPP
        R.....K.
    };
    board.set_turn(Color::Black);
    board.lose_castle_rights(ALL_CASTLE_RIGHTS);

    // ... tests ...
}

Profiling

Combined with flamegraph, you can get a flamegraph of the engine's performance. For example:

sudo cargo flamegraph --bench pvp_benchmark

The pvp_benchmark simulates a game where the engine plays against itself, producing a realistic cross-section of the engine's typical performance. Flamegraphs visually represent the time spent in different parts of the code, with wider sections indicating more time consumed (relative to other codepaths). This helps identify performance bottlenecks and optimize critical codepaths during development.

Various other benchmarks are available in the benches directory.

Codebase structure

• common contains code that is shared between the engine and the precompiler. This is primarily the Bitboard type.
• precompile contains the precompiler, which generates the ZobristHashTable tables and magic bitboard calculation (see this for background).
• src contains the engine's main logic, including the AlphaBetaSearcher and MoveGenerator.