Reinforcement learning on GB Tetris

1. Obtain a ROM of Tetris for Gameboy, move to the lib directory, and rename tetris.gb
2. In your commandline, navigate to the TetrisLearning directory and install conda environment with with conda env create -p ./.conda environment.yml. You may swap out the -p ./.conda argument with -n some_name if you prefer.
3. Activate the environment: conda activate ./.conda
4. Train the model by running python run/run_training.py
5. [Optional] Track progress of the model training via Tensorboard. While the model is running, open a seperate terminal, navigate to the TetrisLearning directory, and run tensorboard --logdir ./board
6. Once training has finished, play the model using the play_model.ipynb notbook

Resources

• Pan Docs: Low-level details on the Gameboy
• Reverse Engineering the Gameboy Tetrs: Memory locations in GB Tetris
• RAM Map for GB Tetris
• Paper: Human-level control through deep reinforcement learning: The setup for the CNN we use, as well as the frame stacking method are pretty much directly pulled from this paper. See the methodology section for specifics. For a simpler overview, see this related presentation from the authors
• Paper: Asynchronous Methods for Deep Reinforcement Learning: The A2C model used is from this paper.