Installation

You need to have pytorch pre-installed. Easy to use download scripts can be found on their website.

$ git clone https://github.com/jaybutera/tetrisRL
$ cd tetrisRL
$ python setup.py install

or

$ pip install tetrisrl

Layout

• dqn_agent.py - DQN reinforcement learning agent trains on tetris
• supervised_agent.py - The same convolutional model as DQN trains on a dataset of user playthroughs
• user_engine.py - Play tetris and accumulate information as a training set
• run_model.py - Evaluate a saved agent model on a visual game of tetris (i.e.)

$ python run_model.py checkpoint.pth.tar

Usage

Using the Environment

The interface is similar to an OpenAI Gym environment.

Initialize the Tetris RL environment

from engine import TetrisEngine

width, height = 10, 20
env = TetrisEngine(width, height)

Simulation loop

# Reset the environment
obs = env.clear()

while True:
    # Get an action from a theoretical AI agent
    action = agent(obs)

    # Sim step takes action and returns results
    obs, reward, done = env.step(action)

    # Done when game is lost
    if done:
        break

Example Usages

Play Tetris for Training Data

Play games and accumulate a data set for a supervised learning algorithm to trian on. An element of data stores a (state, reward, done, action) tuple for each frame of the game.

You may notice the rules are slightly different than normal Tetris. Specifically, each action you take will result in a corresponding soft drop This is how the AI will play and therefore how the training data must be taken.

To play Tetris:

$ python user_engine.py

Controls:
W: Hard drop (piece falls to the bottom)
A: Shift left
S: Soft drop (piece falls one tile)
D: Shift right
Q: Rotate left
E: Rotate right

At the end of each game, choose whether you want to store the information of that game in the data set. Data accumulates in a local file called 'training_data.npy'.

Example supervised learning agent from data

Run the supervised agent file and specify the standard training data file generated in the previous step as a command line argument.

$ python supervised_agent.py training_data.npy

Example reinforcement learning agent

# Start from a new randomized dqn agent
$ python dqn_agent.py
# Start from a the last recorded dqn checkpoint
$ python dqn_agent.py resume
# Specify a custom checkpoint
$ python dqn_agent.py resume supervised_checkpoint.pth.tar

The DQN agent currently optimizes on a metric of freedom of action. In essence the agent should learn to maximize the entropy of the board. A player in Tetris has the most freedom of action when the area is clear of pieces.

Watch a checkpoint play a game

$ python run_model.py checkpoint.pth.tar