RCVL: Reverse Curriculum Hierarchical Vicinity Learning.

This repository was developed for my master's thesis in Artificial Intelligence. It contains a hierarchical algorithm based on two levels in which the higher level draws the high-level path with milestones (subgoals) and the low hierarchy performs the primitive steps in the sub-trajectories between those milestones.

The algorithm is trained and tested in Simple Minigrid environemt: Empty Room 15x15 and FourRooms 15x15.

Dependencies

• Python 3.7
• PyTorch 1.7.1
• OpenAI Gym 0.17.2
• Gym Simple MiniGrid 2.0.0

Usage

Running experiments:

1. For execution of training please run train_rcvl.py. Make sure to insert job_name and to adjust any of the parameters if needed.
2. For execution of testing please run test_rcvl.py. Make sure to insert checkpoint_name, to add the relevant files into the checkpoints directory, and to adjust any of the parameters if needed. The files default checkpoint_name is the algorithms that were presented in the thesis for four_rooms environment. In addition, in the checkpoint directory, it is possible to find the checkpoint for empty room.

~

Trained RCVL policy

For rendering the test, add --render to the configuration file.

Citation

 @phdthesis{Tahar Tair,
    title={Reverse Curriculum Vicinity Learning},
    url={https://upcommons.upc.edu/handle/2117/371021},
    school={UPC, Computer Science Faculty},
    author={Tahar, Tair},
    year={2022},
    month={July}
}