Project goals

The goal of this project is to train an agent to play at Atari Breakout game.

We used the MinAtar Breakout-v1 environment of the gym library and developed two types of agent : a DQN and a Double DQN agent.

Structure of the project

The project is composed of :

• Two Jupyter Notebooks : one for the agent's training (train_agent.ipynb) and another for the explainability (Explainability.ipynb).
• A folder agents composed of the two classes of agents.
• A folder networks_weights where the weights of the agents are saved.
• A folder games where the gif files of the games played by agents are saved.
• Two scripts (train_agent.py and play_game.py) for training and play games by specifying arguments.

Docker

We can use a Docker image to execute the two scripts. This image is based on the Dockerfile which pulls the latest pytorch image available on the Docker Hub and installs the necessary libraries (gym, minatar...).

For building this image : docker build -t rl-breakout-image .

Execution of the training

We can train an agent in a Docker container by running this command :
docker run --rm -v ${PWD}/games/:/app/games -v ${PWD}/networks_weights/:/app/networks_weights rl-breakout-image train_agent.py ARGS
where ARGS are the arguments we can use in the script.

The arguments are :

• --agent : the agent we train (dqn or ddqn) - by default ddqn.
• --gamma: by default 0.99.
• --lr: the learning rate - by default 1e-3.
• --batch_size: by default 32.
• --epsilon_min: by default 1e-2.
• --epsilon_decay: by default 0.999.
• --memory_size: by default 100 000.
• --max_steps: by default 5000.
• --play_game: boolean to choose if we play a game after the training - by default False.

For example, we can execute :
docker run --rm -v ${PWD}/networks_weights/:/app/networks_weights -v ${PWD}/games/:/app/games rl-breakout-image train_agent.py --agent 'dqn' --nb_episode 1000 --play_game True

Notice: the "game" volume is unnecessary if we don't play a game after the training.

Play a game

We can make the agent play and save the game as a gif by running the following command:
docker run --rm -v ${PWD}/games/:/app/games rl-breakout-image play_game.py ARGS

The arguments are :

• --agent : the agent we use (dqn or ddqn) : the best weights of the agent will be loaded - by default ddqn weights.
• --path_weights: path to specific agent weights - by default None.
• --gif_name: by default game.gif.

For example :
docker run --rm -v ${PWD}/games/:/app/games rl-breakout-image play_game.py --agent 'dqn' --gif_name 'dqn-agent-game1.gif'