Paper: Playing Atari the right way!, ECE6504 Project, 2017

Install virtualenv and creating a new virtual environment:

pip install virtualenv
virtualenv -p /usr/bin/python3 atari

Install dependencies

pip3 install -r requirements.txt

Notes:

• Training the agent to play breakout at a reasonable level took about 80 hours on two p100s. Don't even think about running this on a CPU. I would highly appreciate it if you can submit a pull request that makes training faster (I know some of my methods suck).

• The trained models can easily be used to test the performance of an agent on a CPU.

python3 play_cartpole.py

To change the hyperparameters modify mission_control_cartpole.py.

Note:

• This isn't as computationally demanding as Breakout using frames.

python3 play_breakout.py

To change the hyperparameters modify mission_control_breakout.py.

Note:

• I have included the trained model for Breakout after 14 million episodes. Just explore the Results director for Breakout.
• Change train_model to False and show_ui to True to load the saved model and see the agent in action.

python3 generate_dataset.py

Note:

• You might get some directory nt found errors (Will fix it soon) or just figure it out.

python3 generate_model_skip.py

Note:

• This uses the adversarial action conditioned video prediction model.
• Run generate_model.py to use the architecture from [2].

python3 play_breakout_ram.py

To change the hyperparameters modify mission_control_breakout_ram.py.

Note:

• Each run generates a required tensorboard files under ./Results/<model>/<time_stamp_and_parameters>/Tensorboard directory.
• Use tensorboard --logdir <tensorboard_dir> to look at loss variations, rewards and a whole lot more.
• Windows gives an error when : is used during folder naming (this is produced during the folder creation for each run). I would suggest you to remove the time stamp from folder_name variable in the form_results() function. Or, just dual boot linux!

[1] Human Level Control Through Deep Reinforcement Learning

[2] Action-Conditional Video Prediction using Deep Networks in Atari Games