Deep-Q Snake

A Deep Q-Learning agent that teaches itself to play Snake, built with PyTorch and pygame.

How it works

• Network: Linear_QNet, 11 inputs → 256 hidden (ReLU) → 3 outputs, one Q-value per move.
• State (11 inputs): danger straight / right / left · current heading (4) · food direction (4).
• Actions: go straight, turn right, turn left.
• Rewards: +10 for eating food, −10 for dying (or stalling too long), 0 otherwise.
• Training: epsilon-greedy exploration that decays over the first ~70 games, a replay memory of 100k steps, and a short-memory update every step plus a 1,000-sample batch after each game. Discount γ = 0.9, Adam with lr = 0.001.

Files

File	What it does
game.py	pygame Snake environment (SnakeGameAI)
model.py	Linear_QNet and QTrainer (the Q-learning update)
agent.py	state encoding, action selection, replay memory, training loop
helper.py	training graphs: score, reward distribution, loss
model/model.pth	trained weights (loaded automatically by agent.py)

Run it

pip install -r requirements.txt
python agent.py    # press Q to stop and see the training graphs

game.py uses arial.ttf for the score bar. The font isn't included here because its license doesn't allow redistribution; copy any .ttf font into the project folder as arial.ttf.