Tetris RL Agent

This repo is implementing an RL agent integrated with Tetris engine to try to create a DQN agent able to learn how to play Tetris game.

It is implementing the agents and the controllers mentioned in that thesis

Layout

• tetris_agent.py - It contains the RL agent and manipulate the connection between the DQN and the environment.
• run_agent.py - This is the run file that contains also the most of the hyperparameters.
• replay_buffer.py - Please refer to thesis
• tetris_engine.py - This file contains the main class of Tetris engine.
• controllers.py - This file contains many controllers that help to evaluate the states and the reward functions. For more info check
• models
  • dqn_model.py - It is the NN that receives the data.
  • noisy_layer.py - It is a custom NN layer that is used in DQN NN.
• utils
  • drawing_utils.py - It contains functions to draw and plot the data.
  • tetris_engine_utils.py - It contains the necessary functions for the Tetris Engine.
• dqn_models_stats - It is the directory that store the saved models from the NN.

Usage

Run the run_agent.py file with optional arguments

python run_agent.py

The arguments are

--cuda # To decide on what the model shall be trained. CPU or GPU.
--width # The width of the board the default is 10.
--height # The width of the board the default is 20.
--mode # Whether it is play or train mode. play mode here means to test the NN. 
--model # The name of the model which be loaded to train from or to test it.
--model-dir # The path of the directory where the models will be saved. default=dqn_models_stats.
--play-times # Number of games to be played to test the NN. default= 100.