puppetmaster12 / snakeai Goto Github PK

This project is part of my learning experience on machine learning and deep learning. The project is comprised of a regular snake game built with pygame. In addition an AI agent added using Deep Q Learning built using Pytorch. I wanted to learn how to use reinforcement learning in combination with a simple game like snake and improve on it to make it more robust in solving the snake game manouvering.

I recommend installing Anaconda or venv and setting up a virtual environment before installing the packages

Requirements.txt file contains all the packages. Use pip install -r requirements.txt to install all the packages required for the project.

Running the snakeGame.py file runs the regular Snake game. Use the directional arrow keys to play the game. The game currently terminates if you reach a game over status which is if the snake either hits the boundary of the window or hits itself.

Running the agent.py file starts the training of the QLearning model.

The current AI agent is incapable of converging and falls into the trap of predicting the same move on a loop until it hits itself or the boundary. This maybe a problem with insufficient balance in the reward system.

• Build more levels with additional boundaries and mazes for the snake to navigate before reaching the food.
• Improve the AI to reach the food more efficiently.
• Experiment with other algorithms.

27/10/2022

Added a penalty for 100 moves without food and terminated the frame for 200 moves without food