This repository provides the implementation for the Actor-Attention-Critic (MAAC) algorithm, a multi-agent reinforcement learning approach designed for cooperative tasks. It is targeted at researchers and practitioners in MARL, offering a framework to explore attention mechanisms for improved coordination among agents.

How It Works

MAAC utilizes an actor-critic architecture where each agent has its own actor but shares a centralized critic. The critic employs an attention mechanism to selectively focus on relevant information from other agents, enabling more effective credit assignment and coordination in complex multi-agent scenarios. This attention-based approach aims to overcome limitations of simpler coordination methods by dynamically adapting to the interactions between agents.

Quick Start & Requirements

• Primary install: pip install -r requirements.txt (after cloning the repo)
• Prerequisites: Python 3.6.1+, PyTorch 0.3.0.post4, OpenAI Gym 0.9.4, Tensorboard 0.4.0rc3, Tensorboard-Pytorch 1.0, OpenAI baselines (specific commit hash provided).
• To view options: python main.py --help
• Environments: fullobs_collect_treasure (max episode length 100), multi_speaker_listener (max episode length 25).

Highlighted Details

• Implements the Actor-Attention-Critic (MAAC) algorithm for MARL.
• Features a centralized critic with an attention mechanism for inter-agent communication.
• Includes implementations for "Cooperative Treasure Collection" and "Rover-Tower" environments.

Maintenance & Community

• The code is from a 2019 ICML paper. No recent activity or community links (Discord, Slack, etc.) are mentioned in the README.

Licensing & Compatibility

• The README does not explicitly state a license. The code is provided for research purposes.

Limitations & Caveats

The project relies on older versions of key libraries like PyTorch (0.3.0) and OpenAI Gym (0.9.4), which may present compatibility challenges with current MARL research and development stacks. The lack of recent maintenance or community support could also hinder adoption and troubleshooting.