Dots and Boxes Engine

This repository contains a lightweight Dots-and-Boxes engine along with an AlphaZero-style self-play loop. The implementation supports arbitrary (rows, cols) board sizes and is intended as a foundation for reinforcement learning experiments.

Package Overview

• dots/engine.py: Core game engine, including game state, legal moves, scoring, and observation tensor encoding.
• dots/alphazero.py: MCTS implementation, a random policy/value network stub, and a self-play loop that returns (observation, policy, value) training data.
• dots/__init__.py: Public API exports.

Quick Start

from dots import DotsAndBoxes, RandomPolicyValueNet, self_play

game = DotsAndBoxes(rows=2, cols=3)
network = RandomPolicyValueNet(game.new_game().action_size)
trajectory = self_play(game, network, simulations=25, temperature=1.0)

Command Line Usage

Run a small self-play training loop (defaults: 4x4 board, 50 simulations per move) and save loss plot and best models into runs/:

python -m dots.cli train

Print a single self-play game to the terminal:

python -m dots.cli print-game

Notes

• The RandomPolicyValueNet is a stub meant to be replaced with a learned model.
• The engine uses NumPy for state storage and tensor creation.