morality-gym 0.1.0

Gymnasium-compatible trolley-problem environments for studying reinforcement-learning agents under moral constraints.

Morality Gym Tabular

Morality Gym Tabular provides Gymnasium-compatible trolley-problem environments for studying how reinforcement-learning agents behave under moral constraints.

The repository has two layers:

• morality_gym/: the environment package, morality chains, cost function utilities, and lightweight wrappers.
• baselines/ and omnisafe/: experiment code used for the paper benchmarks. These are intended for repository users, not as the minimal pip-facing API.

Installation

For development from this repository:

git clone https://github.com/SimonRosen173/morality-gym-tabular.git
cd morality-gym-tabular
pip install -e .

The lightweight environment path depends on numpy, matplotlib, and gymnasium. The experiment stack also uses additional packages such as torch, stable-baselines3, omnisafe, pandas, tqdm, and cluster tooling where relevant.

Basic Usage

from morality_gym import make

env, morality_chain = make(
    env_id="SwitchStandard-HumanA-v1",
    morality_chain_id="Utility",
)

obs, info = env.reset(seed=42)

done = False
while not done:
    action = env.action_space.sample()
    obs, reward, terminated, truncated, info = env.step(action)
    done = terminated or truncated

env.close()

Environment IDs use:

{scenario_id}-{variant_id}-v1

Examples include SwitchStandard-HumanA-v1, PushStandard-HumanA-v1, Switch5-Human-v1, and PushOrSwitch-Human-v1.

Safe-RL Example

The standalone example in examples/train_safe_rl_agent.py shows how to:

1. create an environment and morality chain,
2. wrap the environment so info["cost"] is emitted at each step,
3. train a small constrained Q-learning agent with a Lagrangian cost penalty.

Run a short smoke test:

python examples/train_safe_rl_agent.py \
  --episodes 10 \
  --eval-episodes 3 \
  --max-episode-steps 25 \
  --log-every 5

Run a longer example:

python examples/train_safe_rl_agent.py --episodes 500 --eval-episodes 50

This script is a readable demonstration of the cost-wrapper pattern. It is not the paper benchmark implementation.

Experiments

Paper-style experiments are configured and run through the benchmarker under baselines/. In short:

python baselines/cli.py --create exp_p9xe.json
python baselines/cli.py --exec-run p9xe/run_0.json

See baselines/README.md and baselines/benchmarker/README.md for the experiment configuration format, local execution, SLURM execution, logs, and result consolidation notes.

Documentation

Resource	Description
morality_gym/README.md	Supported environment ID syntax and scenarios
examples/	Small scripts for interacting with and evaluating environments
baselines/README.md	Baseline learner and benchmarker overview
baselines/benchmarker/README.md	Experiment configuration and execution details

Citation

If you use this framework in your research, please cite:

@misc{rosen2024moralitygym,
  author = {Rosen, Simon and Singh, Siddarth and Robertson, Helen Sarah and Gelo, Ebenezer and
            Suder, Ibrahim and Nangue Tasse, Geraud and Williams, Victoria and
            James, Steven and Rosman, Benjamin},
  title = {Morality Gym Tabular: A Framework for Moral Dilemmas in Reinforcement Learning},
  year = {2024},
  publisher = {GitHub},
  journal = {GitHub Repository},
  howpublished = {\url{https://github.com/SimonRosen173/morality-gym-tabular}}
}

Contributing

Contributions are welcome. See CONTRIBUTING.md.

License

This project is licensed under the terms of LICENSE.