Snake-AI

This project contains a small reinforcement learning playground built around the classic Snake game. Besides a playable version of the game it includes a snake_env.py environment and a simple Deep Q-Network (DQN) agent used for experiments.

For a step-by-step introduction see TUTORIAL.md.

Features

• Play a classic snake game with animated UI (game.py).
• Gym-like environment (snake_env.py) for training RL agents.
• DQN training script (training.py).
• Example agent weights (dqn_snake_model.pth).
• Script to run the trained model (bench.py).
• Customisable environment via config.json.

Installation

1. Install Python 3.8 or later.
2. Install the project dependencies:

pip install -r requirements.txt

The requirements file installs both pygame for rendering and torch for the DQN implementation.

Running the Game

To play the Snake game with keyboard controls run:

python game.py

The configuration in config.json controls various game options such as obstacle count, movement speed and window size.

Training the AI

The file training.py trains a DQN agent using SnakeEnv. The script reads the same configuration file as the game. Training can be started with:

python training.py --config config.json [--render]

The optional --render flag shows the environment during training. When enabled, the same graphical interface used by bench.py and the regular game is displayed so you can watch the learning progress.

A path to a directory may also be supplied, in which case all JSON files in that directory are used for training runs.

After training finishes a model file dqn_snake_model_<name>.pth is written containing the learned weights.

Running the Trained Agent

A small script bench.py loads a previously saved weight file and lets the agent play using the standard game interface:

python bench.py

This launches the regular game window but actions are chosen by the trained model instead of keyboard input.

Ensure that dqn_snake_model.pth (or your own saved weights) is present in the working directory.

Configuration

config.json holds all tunable parameters. Important options include:

• WINDOW_WIDTH, WINDOW_HEIGHT, CELL_SIZE – dimensions of the game grid.
• USE_STATIC_OBSTACLES, NUM_STATIC_OBSTACLE_CLUSTERS, CLUSTER_MAX_SIZE – control randomly generated obstacles.
• USE_MOVING_OBSTACLES, INITIAL_MOVING_OBSTACLES_COUNT – add moving hazards.
• USE_LIVES, INITIAL_LIVES – optional life system in the manual game.
• BASE_FPS – baseline game speed.

You may duplicate this file under different names to experiment with various settings and then pass the file path using --config when training or playing.

Repository Structure

bench.py      # Run a trained model
config.json   # Default configuration
game.py       # Playable Snake game
snake_env.py  # Gym-like environment used for training
training.py   # DQN training script
textures/     # Sprite images
TUTORIAL.md   # Tutorial for more information

License

MIT – do whatever you like, just keep the notice. 😊