Zero Shot Generalization in Reinforcement Learning from Few Training Contexts

Installation

conda create -n cmdp python=3.12

conda activate cmdp
pip install -r requirements.txt

Experiments

Tabular

To run the tabular experiments, just run

python tabular.py

Context Sample Enhancement (CSE)

To run the CSE experiments from the paper, just run the corresponding script in scripts/. For example, to run simple_dir:

./scripts/simple_dir.sh

By default, the results will be saved in ray_results/ on your local machine. Move the experiments into the experiments/ folder before proceeding to the next step. When this is done, you can run sweeps over the contexts by running the following command:

python sweep.py --env simple_dir --nmesh 50

To plot the results, run the following command:

python plot_evaluation.py --env simple_dir --save_fig

Plots are saved in figures/.

These commands should be run for the following environments:

• simple_dir
• ant_dir
• ant_goal
• cart_goal
• cheetah_vel
• pen_goal