Implements the simplest deep reinforcement learner for Tetris based on the Google DeepMind paper Playing Atari with Deep Reinforcement Learning including

• pyglet
• theano
• lasagne

• Training an agent:
  • Simply type python start_rl.py
  • Hyperparameters can be adjusted in the same file line 71. Please consult the Google DeepMind paper for further information on these parameters:
    • learningrate
    • discountfactor
    • epsilon
    • minepsilong
    • depsilong
  • Withing start_rl.py, the training method can be chosen. See lib/Learning.py for how to instantiate:
    • Deep Q Learner (default)
    • Q-Learner
    • SarsaLearner
    • SarsaLambdaLearner
• Continue training:
  • The script start_rl.py automatically looks for policies of name policy-*.pickle and continues the most recent by default.
• Watch an agent play:
  • Change the variable DRAW within start_rl.py to True.
• Adjust the Tetris environment within start_rl.py and the type of blocks in lib/Shape.py
  • Note that the classical Tetris environment (board size of 14x20 and classical block types) are very difficult to get to convergence as the possibility of scoring is extremely small.

The Tetris framework was originally created by Charles Leifer.

Visit https://github.com/coleifer/tetris

The deep reinforcement learning code by Nathan Sprague inspired the implementation for this project.

Visit https://github.com/spragunr/deep_q_rl