Computational Benefits of Intermediate Rewards for Hierarchical Planning

We provide the code used to run the MiniGrid experiments provided in the paper.

Features

All our custom MiniGrid environments are available in gym-minigrid/gym_minigrid/envs/custom.py

For asynchronous Q-learning:

• Script to train: scripts/qlearn.py
• Script to evaluate: scripts/qlearn_evaluate.py

For Deep RL algorithms (A2C, PPO, DQN):

• Script to train: scripts/train.py
• Script to evaluate, scripts/evaluate.py

See experiments/ folder to run all experiments conducted in the paper. We provide a sample parser file Log_Parser.ipynb to gather results presented in paper (average steps, rewards, win rate) for all seeds.

Installation

1. Clone this repository.

2. Install gym-minigrid environments and torch-ac RL algorithms:

pip3 install -r requirements.txt
cd torch-ac
pip3 install -e .