MA²E : Addressing Partial Observability in Multi-Agent Reinforcement Learning with Masked Auto-Encoder

🎯 Description

This is the official code repository for the paper "MA²E : Addressing Partial Observability in Multi-Agent Reinforcement Learning with Masked Auto-Encoder" accepted by ICLR 2025. MA² is a novel approach to addressing partial observability in multi-agent reinforcement learning (MARL). MA²E enables agents to infer global information solely from local observations by leveraging masked auto-encoders (MAE), eliminating the need for explicit communication.

⚙️ Installation

The code is based on pymarl2. For detailed information, refere to the installation instructions of pymarl2 and SMAC.

1️⃣ Cloning MA²E

git clone https://github.com/cheesebro329/MA2E.git

2️⃣ Donwload and setup StarCraftII

bash install_sc2.sh

3️⃣ Install required packages

pip install -r requirements.txt

🎮 Running Script

🏃Run an experiment

bash run.sh config_name env_config_name map_name_list (arg_list threads_num gpu_list experinments_num)

✔️ Example

bash run.sh qmix sc2 3s_vs_5z use_MT=True 3 0 3

• Run QMIX+MA2E in SMAC 3s_vs_5z scenario

• use_MT means executing the model plugs in MA2E into the baseline algorithm.

📌 Citation

@inproceedings{
kang2025mae,
title={{MA}\${\textasciicircum}2\$E: Addressing Partial Observability in Multi-Agent Reinforcement Learning with Masked Auto-Encoder},
author={Sehyeok Kang and Yongsik Lee and Gahee Kim and Song Chong and Se-Young Yun},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=klpdEThT8q}
}