This is a repository for developing a Rainbow DQN. The starting seed is a miniproject from Udacity's Nanodegree in Reinforcement Learning. Implemented:
-
Deep Q-Network with memory replay for learning (DQN)
-
Double DQN (for some reason it still does not learn)
-
Duel (Double) DQN (when not using DDQN it learns well, although a bit slower than the DQN)
-
Importance sampling
The architecture of the network can be specified when instantiating the agent itself with a list of number of units (for fully-connected layers).
To see the application in another game: Picking up bananas
Python files Model and dqn_agent contain the definition of the model and the agents. The jupyter notebook shows how to use and train them
-
Numpy
-
Matplotlib
-
OpenAI Gym (LunarLander-v2)
-
Pytorch
-
pyvirtualdisplay
-
box2d
-
xvfb