Made Snake game and trained computer to play it itself using g NeuroEvolution of Augmenting Topologies Genetic Algorithm

The ai gets 8 inputs from the game. The inputs are the distance from the snake to the 3 walls front left and right, the distance from head to snake body in three directions as well distance to the food in x and y.

The inputs are then fed to the neural network. The output of the neural network is the direction (front,left,right) the snake should move in.The direction is chosen by taking the direction with the highest output value.

The fitness function is the score of the snake. It is calculated using the following rules:

+1 if the snake gets closer to the food

-1 if the snake gets further from the food

+100 if the snake eats the food

-100 if the snake dies from starving

+0.1 for every frame the snake is alive

stop the game if the snake dies

The neural network is trained using the neat algorithm. The neat algorithm is a genetic algorithm that evolves the neural network to play the game. The neat algorithm is implemented in the neat-python library. The neat algorithm is explained in the neat-python documentation.

The visualization of the neural network is done using graphviz (use the runcheck.py file to visualize checkpoints and winner). One can visualize all checkpoints and winners after training is done.

Clone the repo and the submodules

git clone --recurse-submodules https://github.com/om-khandwala/SnakeAI.git

Install the requirements

pip install -r requirements.txt

Train the game

python ai.py

You can change settings in the ai.py file.

RUNS_PER_GAME=10
CORE_COUNT = multiprocessing.cpu_count()-4
GAME_SIZE=(200,200)
FOOD_TIMER=(max(GAME_SIZE[0],GAME_SIZE[1])//10) * 1.414

in the runcheck.py file.

CHECKPOINT_DIR = './checkpoint'
CONFIG = 'config'
GAME_SIZE=(200,200)
FOOD_TIMER=(max(GAME_SIZE[0],GAME_SIZE[1])//10) * 1.414
RUNWINNER=False
GUI=True
GRAPHS=False
SNAKE_SPEED=10