This is a project to train a neural network to play the game snake. The neural network is trained using a genetic algorithm. The project is written in python and uses the arcade library for the game.
pip install -r requirements.txt
- Generate dataset
how to generate dataset
Snake moves automatically with rules (if, else, for, while, etc) and write information in a dataset.csv file with pandas
you can type below command in terminal to run this file:
python main_ai.py
python generate_dataset.py
We implemented the game with simple artificial intelligence that we made with a few if _else conditions so that the snake moves towards the apple without the intervention of the user, and after a few moves and eating the apple, all the movements of the snake, which we named as X, were created using pandas. We save the shape of the data frame.
Null values prevent the execution of the code and reduce the accuracy of the algorithm, we inevitably remove these values with the code in the preprocess section
- Train naural network
you can type below command in terminal to run this file:
python train.py
We read the dataset with pandas and convert it to numpy array and create a neural network using tensorflow. The input layer is the number of columns of the dataset. The hidden layers are completely optional and also the activation function. Of course, the last layer, because our problem is classification, due to the number of outputs, which is four keys, means the answer to the directions of the snake: up and down, left and right, so it is better for the last layer function to be activated by softmax, it is not a rule based on experience. With scikit learn, we divide all the data into two parts, train and test, in order to test it after learning.
We get loss, accuracy with Tensorflow. I save all the things I said in the train section in the file called snake_game_model.h5.
-
u: up direction
-
r: right direction
-
d: down direction
-
l: left direction
-
w: wall
-
a: apple
-
b: body
For example
wr is the distance between the snake and the right wall
ad=350 means there is an apple at the bottom of the snake, and the distance between the apple and the head of the snake is 350
bl=0 means there is no part of the snake's body on the left side of the snake's head
The direction is the same as label, target or y, which is one of the values u, r, d, l
| loss train | accuracy train |
|---|---|
| 0.0017 | 1.0000 |
| loss test | accuracy test |
|---|---|
| 0.0014 | 1.0000 |
](https://github.com/KhadijehValipour/Snake_AI/blob/main/assents/loss_acc.png)
- Play game
you can type below command in terminal to run this file:
python main_ml.py
I will give the file that I saved with the name snake_game_model.h5 to the game that was implemented with simple artificial intelligence to use it and run it. The picture below is the output of the snake game with a neural network.
python main_manual.py
The output of the execution main_ml.py
