This is the implementation of DQN and Model Free Episodic Control
#Introduction This package provides DQN and Episodic Control. The DQN implementation is based on spragunr/deep_q_rl and the Episodic Control is written by myself.
Model Free Episodic Control, C. Blundell et al., arXiv, 2016.
Human-level control through deep reinforcement learning, V. Mnih et al., Nature, 2015.
I have contacted the author C. Blundell of Model Free Episodic Control. He told me he was using approximate KNN to speed up episodic control, however, he did not tell me details. So I used annoy to do KNN, and I rebuild the search tree frequently.
related repo:https://github.com/astier/model-free-episodic-control
#Dependencies
Game roms should be stored in directory roms which stays next to dqn_ep.
Model-Free-Episodic-Control
├ dqn_ep -> source codes
├ roms -> game roms
└ README.md
###Tips: I made some changes to DQN so that we do not need OpenCV any more. In addition if your python has OpenAI gym then you do not need to install Arcade Learning Environment (https://github.com/mgbellemare/Arcade-Learning-Environment) Check spragunr/deep_q_rl's readMe to know more about how to install ALE.
Personally, I recommend using OpenAI gym because it not only can be installed by simply pip install gym
but also provides us atari game roms (For instance /Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/atari_py/atari_roms
on my mac).
###Dependencies for running DQN
Arcade Learning Environment or OpenAI gym
Numpy and SciPy
A reasonable GPU
###Dependencies for running Episodic Control
Arcade Learning Environment or OpenAI gym
Numpy and SciPy
annoy for approximate KNN
A reasonable CPU
examples:
THEANO_FLAGS='device=gpu0, floatX=float32' python run_nature.py
THEANO_FLAGS='device=gpu1, floatX=float32' python run_nature.py -r ms_pacman
python run_episodic_control.py
To get more running details, we can use python run_episodic_control.py -h
or python run_nature.py -h