Connect Four AI

An AlphaZero implementation for playing Connect Four with a web interface.

Installation

1. Clone the repository:

git clone https://github.com/PaulTbbr/connect_4.git
cd connect_4

2. Install dependencies:

pip install -r requirements.txt

Training

To train AlphaZero on connect 4:

python -m src.main train --game connect_four --n_blocks 9 --n_filters 128 --lr 0.001 --weight_decay 0.0001 --c 2 --n_searches 600 --n_iterations 8 --n_self_play_iterations 500 --n_epochs 4 --batch_size 128 --temperature 1.25 --dirichlet_epsilon 0.25 --dirichlet_alpha 0.3 --checkpoints_dir checkpoints/ 

For faster training with parallel self-play, add the flag --parallel.

Key Parameters:

• --n_parallel_games: Number of concurrent self-play games (typically 50-200)
• --n_self_play_iterations: Total games per iteration (should be multiple of n_parallel_games)
• --batch_size: Mini-batch size for NN training
• --c: Exploration constant (higher = more exploration)
• --temperature: Controls policy randomness (1.0 = proportional to visit count)

Playing

Run the web application with:

python -m webapp.app

The game is hosted at http://localhost:5000.

Game Modes

• Human vs Human: Two-player mode
• Human vs MCTS: Play against Monte Carlo Tree Search
• Human vs Alpha-MCTS: Play against an AlphaZero-trained model

Credits

Based on the AlphaZero implementation from:

• Tutorial: https://www.youtube.com/watch?v=wuSQpLinRB4&list=WL&index=20
• Repository: https://github.com/foersterrobert/AlphaZeroFromScratch