Genetic algorithm with neural network to play the snake game

To start the evolution. This will generate weights.txt and bias.txt files

python3 evolve.py

To play the game with the model already evolved. This will load weights.txt and bias.txt files

python3 play.py

The model is a simple neural network

Input: 9 nodes
Hidden layer: 6 nodes
Output: 3 nodes

The input of the neural network is the snake vision in 3 directions (relative to the snake): left, front, right. Each vision is composed of 3 values:

- 0/1 flag indicating if there is an obstacle in the next square
- 0/1 flag indicating if there is food in sight
- 0..1 value indicating the normalized distance to the tail

Indicates if the snakes should go left, forward, or right.

This is the result for the evolution of 20 generations of 1000 individuals in a game with size 10x10.

At first, I tried with a simplier input, just using this for each direction vision

- 0/1 flag indicating if there is an obstacle in the next square
- 0/1 flag indicating if there is food in sight

But the result of the evolution with that input ended up with a snake that was good at caching food but really bad at not crashing with its tail. So I added the distance to the tail to the input and that, somehow, ended up with a snake with a more twisting behaviour that makes it last longer in the game.

There is a lot of room for improvement but I don't want to spend more time on this UwU