This was my project for Project Week Fall 2020 at The Putney School.
I'm using tf_agents to build a neural network using DDQN to solve the game 2048.
If you want to see an agent of mine in action, use visualisation.py.
It's also a good place to start if you want to look around the code.
I think it's pretty well commented.
env.py contains the game logic as a TensorFlow PyEnvironment.
agent.py is for creating and training agents.
pg_implementation.py is a graphical interface for 2048 playable by both humans and robots, implemented in pygame.
visualisaton.py uses pg_implementation.py to show agents playing the game.
data_plots.py can make nice plots of data collected during training.
You control the pygame interface like this:
Move with arrow keys or WASD.
Press r to restart.
Press b to turn the bot off and on.
This project depends on the following packages:
- tensorflow
- tf_agents
- numpy
- pygame (only
pg_implementation.py) - matplotlib (only
data_plots.py) - scipy (only
data_plots.py)
Feel free to send me any questions at [email protected]
2048 was originally created by Gabriele Cirulli and can be found here.
This software is published under the MIT License, see LICENSE.txt.