繁體中文 | English | 日本語

Using the PPO algorithm combined with the CNN architecture, we achieve AI agent control of the Snake game.

The basic structure of the game is located in game.py, and wapper.py is used to package the environment for model operation.
The code for training and testing the model is located in train.py and test.py.
The final trained model is in the model/ folder, while models during training are in the cnn_log/log/ folder.
cnn_log/logs/PPO_0/ contains tensorboard logs of the training process.

On a $12×12$ board:

• After 50 million steps: the average length of the snake reaches 74
• After 100 million steps: the average length of the snake reaches 98

│
├─model
├─cnn_log
│  ├──log
│  └──logs/PPO_0
│ game.py
│ test.py
│ train.py
│ wrapper.py
│ snake.yml
│

Developed using Python in an Anaconda environment.

conda create --name Snake python=3.10.14
conda activate Snake

conda install pytorch=2.2.2 torchvision pytorch-cuda -c pytorch -c nvidia
conda install tensorflow
pip install pygame==2.5.2 gym==0.26.2 gymnasium==0.29.1 stable-baselines3==2.3.2 tensorboard==2.9.1

or use the packaged environment snake.yml

conda env create --file snake.yml --name Snake

Inspired by this Video.
Thanks for the explanation and analysis of many knowledge points.