Demonstration of Monte Carlo Tree Search

The MCTS engine is built in rust.

Running docker compose up --build wasm-builder should compile the source code and produce webassembly files in /docs/pkg

The UI is built in ractive/bulma, so javascript handles user interaction and display, and rust/wasm handles high performance computation.

Two games are provided: connect 4 and tictactoe. The MCTS algorithm itself has no knowledge of the game mechanics, it can learn to play ANY game that implements the GameState trait:

pub trait GameState {
    type CommandType: Copy + Debug;

    fn player(&self) -> Player;
    fn result(&self) -> Result;
    fn commands(&self) -> &Vec<Self::CommandType>;
    fn apply(&self, command: &Self::CommandType) -> Self;
}

Build

docker compose up --build wasm-builder

Compiles rust source into web assembly, makes available to statit site

Deploy locally

Just host the files, e.g.

python3 -m http.server --directory docs

Performance Notes:

Some rough performance sampling:

x86 VirtualBox

100000 playouts in 470.533265ms seconds.

• 212,539/sec

Chrome/WebAssembly

Ran 100000 playouts in 0.727 seconds

• 137,551/sec

Firefox/WebAssembly

Ran 100000 playouts in 0.857 seconds

• 116,686/sec

Python / Numpy

Ran 1000 playouts in 1.58 seconds

• 632/sec

Note that the Rust version still isn't doing the obvious optimization of reusing nodes between moves, although this doesn't affect number of playouts, but still runs more than 100x faster than my previous Python algorithm.