WARNING : This projet is a work in progress, onging changes are likely to break some of the functionalities

CybMASDE

CybMASDE (Cyber Multi-Agent System Design Environment) is a modular and extensible platform for the design, training, analysis, and deployment of multi-agent systems (MAS). It implements the MAMAD method (MOISE+MARL Assisted MAS Design), combining:

• Modelling: generation of simulated environments (via World Models or MCAS)
• Training: multi-agent learning constrained by organizational specifications (MOISE+MARL)
• Analyzing: inference and explainability of organizational structures (Auto-TEMM / TEMM)
• Transferring: deployment and synchronization with real environments

Its purpose is to produce joint policies that are efficient, stable, and explainable, applicable both in academic research and industrial contexts (with a strong focus on cyber-defense).

---

🚀 Key Features

• Structured project creation and validation (init, validate)

• Automatic environment modelling via World Models (VAE, LSTM, JOPM) or manual modelling via MCAS

• Multi-agent training with MARLlib + Ray RLlib (MAPPO, MADDPG, QMIX, etc.)

• Native integration of MOISE+MARL organizational constraints (action masking, reward shaping)

• Organizational analysis with Auto-TEMM: trajectory clustering, role/goal extraction, organizational metrics (SOF, FOF, OF)

• Automatic or manual refinement loops combining training and analysis

• Deployment into real environments, with two execution modes:

  • DIRECT: policy embedded in the agents
  • REMOTE: policy executed by CybMASDE, actions sent to agents via API

• Result tracking and export (logs, metrics, visualizations, organizational specifications in JSON/CSV/YAML)

---

🏗 Software Architecture

CybMASDE is organized around a Python backend orchestrated by a REST API, accessible via:

• a unified CLI (cybmasde ...) for automation (batch/HPC),
• an Angular-based GUI for project configuration and visual monitoring.

Core technologies:

• Python 3.8+ / Flask: backend and REST API
• PyTorch: training World Models (VAE, LSTM)
• MARLlib + Ray RLlib: scalable multi-agent reinforcement learning
• Optuna: hyperparameter optimization
• Angular: modern and ergonomic GUI

---

⚙️ Installation

git clone https://github.com/julien6/CybMASDE.git
cd CybMASDE
pip install -r requirements.txt

Optional: for the GUI

cd frontend
npm install
npm start

---

📘 Quickstart

Create and configure a project

cybmasde init -n overcooked_test --template worldmodel
cybmasde validate

Run the full pipeline in fully automated mode

cybmasde run --full-auto \
  --project ./overcooked_test \
  --max-refine 10 \
  --reward-threshold 3.5 \
  --std-threshold 0.05 \
  --accept-inferred

Deploy a policy into a real environment

cybmasde deploy --remote --api http://localhost:8080/api

Export results

cybmasde export --format json --output ./results

---

📂 Project Structure

<project_name>/
│── project_configuration.json   # main configuration
│── modelling/                   # simulated environments + MCAS
│── training/                    # hyperparameters, checkpoints
│── analyzing/                   # metrics, visualizations, inferred specs
│── transferring/                # configuration and deployment
└── label_manager.py             # observation/action label mapping

---

📑 Use Cases

• Academic research: rapid prototyping of MARL environments, organizational explainability via Auto-TEMM
• Cyber-defense: integration with real infrastructures (via environment APIs) to evaluate resilience and adaptability
• Industry: deployment of robust and interpretable policies in distributed environments

---

📖 References

• Method MAMAD: MOISE+MARL Assisted MAS Design
• Organizational framework MOISE+MARL API (MMA): github.com/julien6/MOISE-MARL
• Libraries: MARLlib, Ray RLlib, Optuna
• Associated PhD thesis: see /docs

---

✍️ Author

Developed in the context of a PhD in Distributed AI for Cyber-Defense, by Julien Soule.