Friction-MARL

Multi-Agent Reinforcement Learning with Friction Dynamics

Code companion to: Axiom of Consent: Friction Dynamics in Multi-Agent Coordination (DAI-2601)

Overview

Multi-agent reinforcement learning simulation for testing coordination friction frameworks. The codebase simulates a multi-agent resource allocation environment and evaluates whether coordination failure correlates with the theoretical friction function:

F = sigma * (1 + epsilon) / (1 + alpha)

It runs a 5x5x5 factorial design over alignment (alpha), stakes (sigma), and observation entropy (epsilon), with replications per condition, trains independent Q-learning agents, and computes friction proxies plus regression analyses.

Quickstart

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

python run_experiments.py --output-dir ./results

CLI

python run_experiments.py \
  --n-agents 4 \
  --n-resources 3 \
  --n-replications 30 \
  --n-episodes 10000 \
  --output-dir ./results \
  --seed 123

Outputs: Raw results CSV in results/, analysis tables and plots in results/analysis/.

Default settings are computationally heavy (125 conditions x 30 replications x 10,000 episodes). Reduce episodes or replications for quick tests.

Project Layout

friction-marl/
├── friction_marl/
│   ├── envs/          # Resource allocation environments
│   ├── agents/        # Q-learning agent implementations
│   ├── experiments/   # Experiment configurations
│   └── utils/         # Analysis and plotting utilities
├── run_experiments.py # Main entry point
├── requirements.txt   # Dependencies
└── pyproject.toml     # Package configuration

Related Papers

• Axiom of Consent (DAI-2601): arXiv:2601.06692
• ROM (DAI-2503): arXiv:2601.06363

Authors

• Murad Farzulla -- Dissensus AI
  • ORCID: 0009-0002-7164-8704

License

MIT License