Categorical DQN.

Implementation of the Categorical DQN as described in A distributional Perspective on Reinforcement Learning.

Thanks to @tudor-berariu for optimisation and training tricks and for catching two nasty bugs.

Dependencies

You can take a look in the env export file for the full list of dependencies.

Install the game of Catch:

git clone https://github.com/floringogianu/gym_fast_envs
cd gym_fast_envs

pip install -r requirements.txt
pip install -e .

Install visdom for reporting: pip install visdom.

Training

First start the visdom server: python -m visdom.server. If you don't want to install or use visdom make sure you deactivate the display_plots option in the configs.

Train the Categorical DQN with python main.py -cf configs/catch_categorical.yaml.

Train a DQN baseline with python main.py -cf configs/catch_dqn.yaml.

To Do

• [x] Migrate to Pytorch 0.2.0. Breaks compatibility with 0.1.12.

• [x] Add some training curves.

• [x] Run on Atari.

• [x] Add proper evaluation.

Results

First row is with batch size of 64, the second with 32. Will run on more seeds and average for a better comparison. Working on adding Atari results.