PySnake is an artificial intelligence that learns to control a snake. This AI learns to recognize its environment and choose the best direction to survive. This model uses Genetic Algorithm theory, as there is no training data available.

PySnake Interface

The game is simple. The player / AI control a snake in four directions : UP, RIGHT, DOWN, LEFT. The goal is to eat as many apples as possible, increasing the snake's length. The snake dies when it eats itself or crash in a wall.

SAGA is a distributed memory system which consists of 244 dual/quad socket nodes, interconnected with a high-bandwidth low-latency InfiniBand network. The model has been trained on SAGA servers with an access of 16GB ram and 8 nodes, through 1000 generations of 1500 snakes each.

To install, clone this repository:

$ git clone https://github.com/arthurdjn/pysnake

Note that pysnake requires pygame to visualize the game. I you installed pysnake through pip pip install ., it will automatically download all required dependencies. Otherwise, use:

$ pip install pygame
$ pip install numpy
$ pip install json

The application settings are available from the config.ini file. You can create a custom one, as long as you have the necessary parameters.

You can play and launch the application by running the following command:

$ python pysnake

This will call the pysnake/__main__.py file and execute your command. If you choose to render the game, you should see a pop-up window like this:

To play the game, use :

• UP ARROW to move up,
• LEFT ARROW to move left,
• DOWN ARROW to move down,
• RIGHT ARROW to move right,
• SPACE BAR to pause the game,
• r to restart the game,
• v to render the snake's vision,
• g to display the game grid

You can turn it off with render = False in the config.ini file.

There are more key arguments that you can used to launch the application:

• --mode: run either a playble or trainable game,
• --config: read a custom config file,
• --snake: run a snake from a saved file. The snake will play automatically in an unseen environment,
• --replay: run a snake from a saved file, in the same environment it was originally saved,
• --population: train snakes from an existing population.

By default, the game mode used is play. If you just want to play the game, you can explicitly use in addition the --mode key argument:

$ python pysnake --mode play

You can specify --mode train to train snakes.

If you created a custom config.ini, specify it each time you run pysnake:

$ python pysnake --config path/to/your/config.ini

Snakes and generations can be saved. Change the default parameters in the config.ini files. Even though saving the best snakes for X generations depends on your needs, I recommend that you save some of your generations, in case your computer / server shuts down. Saving generations with a saving_steps = 50 prevents you to starting the simulation from zero if something (bad) happens.

Once snakes are saved, load them with one of these commands:

$ python pysnake --snake pysnake/snake.json

$ python pysnake --replay pysnake/snake.json

$ python pysnake --mode train --population pysnake/saves/generation_1500

PySnake architecture depends on 3 elements:

• snake's vision,
• neural network,
• genetic algorithm.

The snake's vision is composed of rays (lines) going from its head to the game borders. These rays are equally spaced, the range depends on the vision_mode parameters (goes from 1 to +infinity). The vision can be visualized as a LiDAR sensor.

8-mode vision

I started this project when I saw the demonstration from Chrisspresso. I read his project page and then wanted to create my own version. The architecture is somewhat similar, but the vision and game are coded differently. In addition, I added a web visualizer so anyone can play with the project 😃 !