This project aims to develop an AI agent that can learn to play the classic game Snake using deep reinforcement learning. The application includes a graphical interface where the interactions of the agent with the game environment can be seen. Closing the graphical interface will show a graph summary of the agent's performance throughout the session.
The environment is a grid where each tile is either empty space, score, or a part of the snake.
The goal of the agent is to collide with as many score tiles as possible while avoiding
collision with snake tiles (itself) and the edges.
With every step of the game the agent is given the following insights about the environment:
- Whether there is a collision risk to the left, right and in front of the head of the snake.
- The direction where the snake is facing.
- The location of the score relative to the head of the snake.
The agent begins by taking random actions to explore the environment and, as it gathers enough experience, it slowly turns to relying on a Deep Q-Learning model to take more calculated actions.
As the only information the agent receives about obstacles in the environment consists of three tiles around the head of the snake the agent is more likely to spiral into itself.
This project was developed with Python 3.10.0 though a different version may work just fine.
Libraries
- PyTorch and NumPy: for the implementation of the agent
- Pygame: for the implementation of the environment
- Matplotlib: for the graph summary of the agent's performance
To run this application first download the source code from the
releases section of
this repository.
Next you might need to install the necessary dependencies as specified under the
requirements section above.
pip install -r requirements.txt
To run the agent simply execute the SnakeAI.sh file or the SnakeAI.bat file included in the
release. It may take several minutes for the agent to show significant improvement.
This release also include a playable version of the game. That is what the Snake.sh and Snake.bat
files are for.
At last, to remove the dependencies:
pip uninstall -y -r requirements.txt