EPyMARL is an extended Python framework for Multi-Agent Reinforcement Learning (MARL), building upon PyMARL. It offers enhanced algorithm support, broader environment integration, and improved training flexibility for researchers and practitioners in MARL.

How It Works

EPyMARL extends PyMARL by integrating new algorithms (IA2C, IPPO, MADDPG, MAA2C, MAPPO) and supporting parameter-free sharing. It adopts the Gymnasium API for compatibility with modern environments and introduces native support for PettingZoo, VMAS, SMACv2, and SMAClite, alongside matrix games. A key feature is the ability to train in general-sum reward environments, moving beyond the common-reward assumption of its predecessor.

Quick Start & Requirements

• Install dependencies: pip install -r requirements.txt
• Install environments: pip install -r env_requirements.txt
• SMAC/SMACv2 require a StarCraft II installation.
• PAC algorithm requires additional dependencies: pip install -r pac_requirements.txt
• Documentation: https://github.com/uoe-agents/epymarl

Highlighted Details

• Supports individual rewards for agents in general-sum environments.
• Integrates logging with Weights and Biases (W&B).
• Includes a plotting script for visualizing experiment results.
• Offers flexibility with implementation details like reward standardization and update strategies.

Maintenance & Community

The project is associated with the University of Edinburgh's agents research group. Updates include migration to Gymnasium and addition of new algorithms and environments.

Licensing & Compatibility

Licensed under the Apache License v2.0. This license permits commercial use and linking with closed-source projects.

Limitations & Caveats

The legacy version (v1.0.0) is required for compatibility with the deprecated OpenAI Gym version 0.21. Some environments, like SMAC and SMACv2, have specific installation prerequisites beyond the main requirements.