mcts-lib 0.3.2

A small and simple library for Monte Carlo tree search
Documentation

mcts-lib

A small and simple library for Monte Carlo tree search.

crates.io docs.rs

This library provides a generic implementation of the Monte Carlo Tree Search (MCTS) algorithm in Rust. MCTS is a powerful heuristic search algorithm for decision-making processes, particularly in games. This library is designed to be flexible and easy to integrate with various turn-based games.

Features

  • Generic implementation of the MCTS algorithm.
  • Flexible Board trait for easy integration with your own games.
  • Includes an example implementation for Tic-Tac-Toe.
  • Alpha-beta pruning for optimization.

Getting Started

Usage

To use this library, you need to implement the Board trait for your game's state representation. Here's a high-level overview of the steps:

  1. Define your game state: Create a struct or enum to represent your game's state.
  2. Implement the Board trait: Implement the Board trait for your game state. This involves defining the logic for:
    • Getting the current player.
    • Determining the game's outcome (win, lose, draw, in-progress).
    • Listing available moves.
    • Applying a move to the board.
  3. Configure MonteCarloTreeSearch: Use the MonteCarloTreeSearch::builder() to create and configure an instance of the search algorithm.
  4. Run the search: Use iterate_n_times to run the MCTS algorithm.
  5. Get the best move: Use get_most_perspective_move to get the best move found by the algorithm.

Example: Tic-Tac-Toe

The library includes a Tic-Tac-Toe implementation that you can use as a reference. See examples/tic_tac_toe.rs.

use mcts_lib::boards::tic_tac_toe::TicTacToeBoard;
use mcts_lib::mcts::MonteCarloTreeSearch;
use mcts_lib::random::CustomNumberGenerator;

// Create a new Tic-Tac-Toe board
let board = TicTacToeBoard::default();

// Create a new MCTS search instance
let mut mcts = MonteCarloTreeSearch::builder(board)
    .with_alpha_beta_pruning(false)
    .with_random_generator(CustomNumberGenerator::default())
    .build();

// Run the search for 20,000 iterations
mcts.iterate_n_times(20000);

// Print the chances
let root = mcts.get_root();
for node in root.children() {
    println!(
        "Move: {:?} = {:.2?}%",
        node.value().prev_move,
        node.value().wins_rate() * 100.0
    );
}

// Get the most promising move
let best_move_node = root.get_best_child().unwrap();
let best_move = best_move_node.value().prev_move;

println!("The best move is: {:?}", best_move);

Building and Testing

  • Build: cargo build
  • Test: cargo test

License

This project is licensed under the MIT License - see the LICENSE file for details.