Skip to content

Introduction

CybMASDE (Cyber Multi-Agent System Design Environment) is a unified research and development platform for designing, training, analyzing, and deploying intelligent multi-agent systems (MAS).
It was originally developed as part of a doctoral research project between Université Grenoble Alpes (LCIS Laboratory) and Thales LAS - La Ruche (Rennes, France), within the context of cyber-defense and autonomous system coordination.

CybMASDE provides both theoretical grounding and practical tooling to bridge the gap between simulation-based AI research and operational, explainable, and deployable multi-agent systems.

🎯 Objectives

The main objective of CybMASDE is to provide a methodological and software framework that supports the entire lifecycle of a multi-agent system (from conceptual modeling to real-world deployment).
It integrates multiple AI paradigms under a unified structure called MAMAD:

MAMAD (MOISE+MARL Assisted MAS Design)

This framework allows:

  • The modeling of environments, world models, and organizations,
  • The training of agent policies using organizationally constrained MARL,
  • The analysis of learned behaviors and emergent dynamics (e.g., via explainability tools like TEMM),
  • The transfer of these behaviors into real or hybrid infrastructures,
  • And the iterative refinement of agent-organization configurations through feedback cycles.

🧩 Theoretical Foundations

CybMASDE is built on two key scientific paradigms:

  1. Multi-Agent Reinforcement Learning (MARL)
    Provides the adaptive intelligence layer where agents learn optimal behaviors via trial and error in cooperative, competitive, or mixed environments.

  2. MOISE+ Organizational Modeling
    Defines explicit organizational roles, missions, and constraints that guide agents’ decisions, ensuring structured autonomy and coordinated behavior.

By coupling MOISE+ and MARL, CybMASDE enables a new research direction called MOISE+MARL, where learning is guided by organizational norms and organizational structures evolve based on learned behaviors.

🧠 Why CybMASDE?

Traditional multi-agent frameworks focus either on:

  • Engineering tools (simulation engines, orchestration layers), or
  • AI libraries (learning algorithms, policy optimization).

CybMASDE combines both worlds, allowing users to:

  • Model complex, multi-level environments (agents, organizations, infrastructures),
  • Train adaptive agents under explicit organizational constraints,
  • Analyze their behaviors through automated and explainable methods,
  • Transfer policies to real-world cyber-physical systems (via APIs or embedded agents),
  • And Refine the full system iteratively.

This makes CybMASDE a comprehensive MAMAD-compliant pipeline from conception to deployment.

🧮 Key Components

CybMASDE is composed of five core modules:

Module Purpose Example Features
Modeling Build or generate environments and world models. Handcrafted environments, latent dynamics models (VAE, RNN, etc.)
Training Optimize agent behaviors under MARL and MOISE+ constraints. MAPPO, MADDPG, QMIX, ROMA, organizational reward shaping
Analyzing Interpret, visualize, and explain agent behaviors. Auto-TEMM, trajectory clustering, explainability metrics
Transferring Deploy learned policies in real or hybrid environments. REST API deployment, trajectory synchronization
Refining Iterate based on analysis outcomes. Feedback loops, re-training triggers, organizational adaptation

Each component is fully configurable via a project configuration file ( project_configuration.json ), making the pipeline flexible for both research and applied contexts.

💡 Typical Use Cases

CybMASDE is designed to support a wide range of application domains:

  • Cyber-Defense: autonomous intrusion detection and response coordination
  • Swarm Robotics: decentralized control and collective adaptation
  • Industrial IoT: resource optimization in distributed networks
  • Microservice Management: adaptive orchestration under changing workloads
  • Research & Simulation: experimentation in MARL and multi-agent organization learning

⚙️ Interfaces and Integration

CybMASDE can be used through several interfaces:

  • CLI (Command Line Interface): for automation, HPC batch runs, and reproducible experiments.
  • Python Library: for integration into research workflows and Jupyter notebooks.
  • Web GUI (Angular): for visual project creation, configuration, and monitoring.

All three interfaces rely on a shared backend API and consistent configuration schema, ensuring that projects remain interoperable across use modes.

🔬 Research Impact

CybMASDE provides a testbed for studying key research questions in distributed AI:

  • How can organizations guide learning in multi-agent systems?
  • How can explainability and interpretability be integrated into MARL?
  • How can policies trained in simulation be safely transferred to real infrastructures?
  • How can agent coordination be maintained under uncertainty or partial observability?

The platform has been validated in multiple experimental scenarios:

  • Company Infrastructure (Cyber-Defense)
  • Drone Swarm
  • Microservices on Kubernetes
  • Warehouse Management
  • Overcooked-AI Environments
  • ...

🧭 Learn More

If you’re new to CybMASDE, start with: