This repository provides the Multi-Agent Particle Environment (MPE), a suite of simulated worlds for researching multi-agent reinforcement learning in mixed cooperative-competitive scenarios. It's designed for researchers and practitioners in MARL, offering a flexible framework to define agent capabilities, interactions, and reward structures.

How It Works

The MPE simulates a 2D world with agents and landmarks, featuring continuous observation and discrete action spaces. Scenarios are defined by implementing make_world, reset_world, reward, and observation functions. This modular design allows for easy creation of new environments and experimentation with different agent communication, movement, and sensing capabilities.

Quick Start & Requirements

• Install: pip install -e . in the root directory.
• Dependencies: Python (3.5.4), OpenAI gym (0.10.5), numpy (1.14.5), pyglet (1.5.27).
• Interactive demo: bin/interactive.py --scenario simple.py

Highlighted Details

• Supports various scenarios including physical deception, covert communication, cooperative navigation, and predator-prey dynamics.
• Environments are structured to facilitate research on communication protocols and emergent behaviors in multi-agent systems.
• Includes a rendering module for visualizing agent interactions.

Maintenance & Community

This project is archived and no longer actively maintained. A maintained fork with numerous fixes, comprehensive documentation, pip installation, and support for current Python versions is available in PettingZoo.

Licensing & Compatibility

The repository does not explicitly state a license. However, it is associated with OpenAI and the original paper, suggesting a research-oriented, potentially permissive license. Compatibility with commercial or closed-source projects is not specified.

Limitations & Caveats

The project is archived and no longer updated, with dependencies on older versions of Python and libraries like OpenAI gym. The lack of explicit licensing may pose challenges for commercial use.