EthicaAI: Emergent Morality via Meta-Ranking 🧠⚖️

"Reason is not just a slave of the passions, but a sovereign that can choose between them." — Amartya Sen

EthicaAI is the official implementation of the paper "Computational Verification of Amartya Sen's Optimal Rationality via Multi-Agent Reinforcement Learning with Meta-Ranking."

This project bridges Moral Philosophy and Multi-Agent Reinforcement Learning (MARL). By formalizing Amartya Sen's theory of Meta-Ranking (preferences over preferences), we demonstrate how AI agents can evolve distinct moral commitments ("Situational Commitment") to solve the Tragedy of the Commons in large-scale social dilemmas.

Fig: Meta-Ranking prevents the "Tragedy of the Commons" at scale (100 Agents). High SVO agents with meta-ranking (blue) sustain resources, while naive agents (gray) collapse.

---

🌟 Key Innovations

1. 🧠 Meta-Ranking Architecture

Unlike traditional methods that treat morality as a fixed parameter (Static SVO), EthicaAI implements a dynamic $\lambda_t$ mechanism that modulates the weight between self-interest and social welfare based on resource abundance.

• Survival Mode: Prioritize self-preservation ($w < w_{survival}$)
• Abundance Mode: Activate moral commitment ($w > w_{boost}$)

2. 📈 Scalability Verified (100 Agents)

We scaled the simulation from 20 to 100 agents, confirming that the emergence of cooperation is robust.

• Super-Linear Inequality Reduction: The mechanism becomes more effective at maintaining fairness as society grows ($f^2$: 5.79 $\to$ 10.2).
• Role Specialization: Emergence of distinct "Cleaner" and "Eater" classes ($p < 0.0001$).

3. 🤝 Human-AI Alignment

We validated our agents against Human Public Goods Game (PGG) data (Zenodo Dataset, 2025).

• Wasserstein Distance < 0.2: Our agents' "Situational Commitment" mirrors human "Conditional Cooperation."

4. 📊 Rigorous Causal Inference

We moved beyond simple correlation.

• HAC Robust Standard Errors: Correcting for temporal autocorrelation.
• Linear Mixed-Effects Models (LMM): Accounting for agent-specific random effects.
• Bootstrap Confidence Intervals: Ensuring statistical solidity.

---

🛠️ Installation

Prerequisites: Python 3.10+, CUDA 12+ (for GPU acceleration).

# 1. Clone the repository
git clone https://github.com/Yesol-Pilot/EthicaAI.git
cd EthicaAI

# 2. Create a virtual environment
python -m venv ethica_env
source ethica_env/bin/activate  # Windows: ethica_env\Scripts\activate

# 3. Install dependencies (JAX, Flax, Statsmodels, etc.)
pip install -r requirements.txt

---

🚀 Usage

1. Run Full Experiment (100 Agents)

Execute the full pipeline including training, evaluation, Causal ATE analysis, and figure generation.

# Run large-scale experiment (Meta-Ranking ON)
python -m simulation.jax.run_full_pipeline large_full

# Run baseline comparison (Meta-Ranking OFF)
python -m simulation.jax.run_full_pipeline large_baseline

2. Run Human-AI Comparison

Verify the alignment between simulation results and human data.

python -m simulation.jax.analysis.human_ai_comparison data/human_pgg.csv simulation/outputs/latest_run/sweep.json

3. Re-generate Figures (Publication Ready)

Generate NeurIPS-style figures (Times New Roman, 300 DPI, PDF/PNG).

python -m simulation.jax.analysis.paper_figures simulation/outputs/latest_run

---

📂 Repository Structure

EthicaAI/
├── simulation/
│   ├── jax/                # Core MAPPO Algorithm & Environment (JAX)
│   │   ├── analysis/       # Statistical Analysis (LMM, Bootstrap, Causal)
│   │   ├── config.py       # Experiment Hyperparameters
│   │   └── run_full_pipeline.py # End-to-End Execution Script
│   └── llm/                # (Experimental) Constitutional AI Prototype
├── submission_neurips/     # LaTeX Sources for NeurIPS 2026
├── figures/                # Generated Figures for Paper
└── requirements.txt        # Python Dependencies

---

📜 Citation

If you use this code or findings, please cite:

@article{heo2026ethicaai,
  title={Computational Verification of Amartya Sen's Optimal Rationality via Multi-Agent Reinforcement Learning with Meta-Ranking},
  author={Heo, Yesol},
  journal={arXiv preprint arXiv:2602.XXXXX},
  year={2026},
  note={Prepared for NeurIPS 2026 Workshop}
}

---

📄 License

This project is licensed under the MIT License. See the LICENSE file for details.

Built with ❤️ by the Antigravity Team.