DamienAllonsius/RL_options

Reinforcement Learning with options

• We train an agent to solve Montezuma's Revenge using the option framework of Sutton-Precup-Singh (see the original article here).
  The code is written with Python 3.6 and uses the ATARI environment from gym.

• To run the script, first install the libraries of requirements.txt and execute python3 main.py.

• To run the experiment on a gridworld environment, clone this repo and, in RL_options folder, clone the repo gridenvs (this gridworld environment is developed by AI-ML team of Universitat Pompeu Fabra (Barcelona)). You can change the shape of the gridworld in gridenvs/example/.

Learning phase

You can change the render of the game by selecting the window and typing: b (Blurred) to switch between a downsampled image and the original image, g (Grayscale) to activate the grayscaling, a (Agent) to switch between the option and the agent view, d (Display) to activate/deactivate the display (of course, the display activation slows down the performances).