TetrisRL

This is a Tetris game played by an AI through reinforcement learning. We use Pytorch for creating the model and OpenCV for displaying the actual game.

Contributors:

David Han, Emmett Breen, Evan Williams

Requirements

• Numpy
• torch
• termcolor
• openCV
• Pillow

Usage

• pip3 install -r requirements.txt
• To run the model, run python3 run.py (a window opens up with the Tetris game using the final trained model)
• For running, run python3 train.py

Other Specifications

There are a few parameters that can be modified while the model is being trained.

• epochs (default = 30000)
• epsilon (default = 1) Used for Epsilon-Greedy Algorithm. Epsilon should never be above 1 and below the epsilon floor (0.001)
• gamma (default = 0.999) Used for Epsilon-Greedy Algorithm which is the rate of decay for epsilon. A gamma closer to 1 will decay at a slower rate, while smaller gamma will decrease epsilon more quickly
• replay_size (default = 100000) Used for keeping track of current states of Tetris
• minibatch_size (default = 200) Parameter for size of minibatch for replay states

Visualization

This is a sample frame of the Tetris game played by the AI: