Dopamine is my favorite research framework of reinforcement learning, for its very readable code and easy to try out new ideas. This project is mainly inspired by dopamine, and can also be said I am learning coding from dopamine.
My design principles are:
- Flexible development: Make it easy to try out research ideas.
- Concentration: Only focus on policy gradient or actor critic style algorithms.
Also inspired by:
sudo apt-get update && sudo apt-get install cmake libopenmpi-dev python3-dev zlib1g-dev
python -m endorphin.atari.train --agent_name=<name of the agent> --base_dir=<name of the base directory> --game_name=<name of the game>
To train a fully-connected network controlling Classic Control using ppo
python -m endorphin.classic.train --agent_name='ppo' --base_dir=/tmp/endorphin/ppo/CartPole --game_name='CartPole-v0'
To train a conv network controlling Atari using a2c
python -m endorphin.atari.train --agent_name='a2c' --base_dir=/tmp/endorphin/a2c/Breakout --game_name='Breakout'
Marc G. Bellemare, Pablo Samuel Castro, Carles Gelada, Saurabh Kumar, Subhodeep Moitra. Dopamine, https://github.com/google/dopamine, 2018.
Mnih et al. Asynchronous Methods for Deep Reinforcement Learning. ICML 2016