/sNNake

Snake game controlled by an artificial neural network

Primary LanguageJava

sNNake

Snake game controlled by an artificial neural network

Quickstart

Once you have cloned the repository,

  1. In command line from project root, run:
cd SnakeBrain
python server.py

  1. Import ./Snake/ into IntelliJ IDEA
    Make sure IS_TRAINING = false in Config.java.

  2. Run Snake.java and watch the snake eat apples by itself.

Training

To train the snake (assuming you have completed quickstart):

  1. Change IS_TRAINING to true in Config.java.

  2. Run Snake.java. Use the arrow keys to move the snake.
    When training is complete, press 'E' to export the training data.

  3. Look in ./TrainingData/ and find the .csv file you want to use to train the snake. Copy this file into ./TrainingData/active/. Make note of the vision type. You are only generating parameters for this vision type. Parameters for other vision types will be unaffected.

  4. Open config.py and change import config_visionX as cfg to the correct vision type. For example, if you wanted to generate parameters for vision1, you would change it to import config_vision1 as cfg.

  5. Run snakecharmer.py (assuming you are already in ./SnakeBrain/):

python snakecharmer.py

Using Training Results

  1. Look in ./Parameters/ for the newly generated .csv file. Copy that to ./Parameters/active/ and rename to weights_visionX.csv, where visionX corresponds to the vision type.

  2. Restart the Python server (server.py)

  3. Make sure VISION_TYPE in Config.java is set to the correct vision type. For example, if you trained for vision1, then set it to Example.VISION_1.

  4. Change IS_TRAINING in Config.java back to false

  5. Run Snake.java and watch your newly trained snake eat apples by itself.

Notes

  • There are total 3 vision types, creatively named vision1, vision2, vision3.

  • Different vision types determine what the snake sees. Play around with them until you find one that works the best. Then let me know (please?)

  • You should know that vision1 sucks. A lot. Unless your board is less than 10x10, don't use it.

  • I'm have not yet determined which is better, vision2 or vision3. They both work pretty well.

  • I'm not a snake master. Thus snakes that learn from me are not masters themselves. The included trained parameters will get your snake to a score of around 30, 40 if you're lucky, and 50 if you're really lucky.

  • If you have trained this snake to do better, please submit a pull request :)

  • The snake's name is Ann. Ann A. Conda. Not Anne. Not Annie. Don't ask why. Just remember what it uses for a brain.

Credits

Special thanks to wollip for creating the snake game this project is based off of.