Implementation of the algorithm described in “No-regret Exploration in Goal-Oriented Reinforcement Learning”, Tarbouriech et al., 2019.

We perform the experiments on two different grid-world gym style environments:

• Maze
• Frozen Lake - a modified version of the OpenAI's enviroement. Here, the game doesn't end when reaching a hole.

Requirements:

• Python 3.5+
• OpenAI Gym
• NumPy
• Pyglet (OpenGL 3D graphics)
• The provided packages

Install the experiment environments by running the following:

cd fozen_lake_env
pip install -e .
cd ..
cd maze_env
pip install -e .

To run the experiment, just open algorithms/uc_ssp/uc_ssp.py and modify the parameters according to the comments at the bottom.