Home

Offline Multi-Agent Reinforcement Learning Datasets and Baselines

Going Off-the-Grid! 🤖 ⚡ 🔌 🔋

Offline MARL holds great promise for real-world applications by utilising static datasets to build decentralised controllers of complex multi-agent systems. However, currently offline MARL lacks a standardised benchmark for measuring meaningful research progress. Off-the-Grid MARL (OG-MARL) fills this gap by providing a diverse suite of datasets with baselines on popular MARL benchmark environments in one place, with a unified API and an easy-to-use set of tools.

OG-MARL forms part of the InstaDeep MARL ecosystem, developed jointly with the open-source community. To join us in these efforts, reach out, raise issues or just 🌟 to stay up to date with the latest developments! 📢 You can contribute to the conversation around OG-MARL in the Discussion tab. Please don't hesitate to leave a comment. We will be happy to reply.

📢 We recently moved our datasets to Hugging Face. This means that previous download links for the datasets may no longer work. Datasets can now be downloaded directly from Hugging Face.

Quickstart 🏎️

Clone this repository.

git clone https://github.com/instadeepai/og-marl.git

Install og-marl and its requirements. We tested og-marl with Python 3.10 and Ubuntu 20.04. Consider using a conda virtual environment.

pip install -e .[tf2_baselines]

Download environment files. We will use SMACv1 in this example. MAMuJoCo installation instructions are included near the bottom of the README.

bash install_environments/smacv1.sh

Download environment requirements.

pip install -r install_environments/requirements/smacv1.txt

Train an offline system. In this example we will run Independent Q-Learning with Conservative Q-Learning (iql+cql). The script will automatically download the neccessary dataset if it is not found locally.

python og_marl/tf2_systems/offline/iql_cql.py task.source=og_marl task.env=smac_v1 task.scenario=3m task.dataset=Good

You can find all offline systems at og_marl/tf2_systems/offline/ and they can be run similarly. Be careful, some systems only work on discrete action space environments and vice versa for continuous action space environments. The config files for systems are found at og_marl/tf2_systems/offline/configs/. We use hydra for our config management. Config defaults can be overwritten as command line arguments like above.

Dataset API 🔌

To quickly start working with a dataset you do not even need to install og-marl. Simply install Flashbax and download a dataset from Hugging Face.

pip install flashbax

Then you should be able to do something like this.

from flashbax.vault import Vault
import jax
import numpy as np

vault = Vault("og_marl/smac_v1/2s3z.vlt", vault_uid="Good")

experience = vault.read().experience

numpy_experience = jax.tree.map(lambda x: np.array(x), experience)

We also provide a simple demonstrative notebook of how to use OG-MARL's dataset API here:

Datasets 🎥

We have generated datasets on a diverse set of popular MARL environments. A list of currently supported environments is included in the table below. It is well known from the single-agent offline RL literature that the quality of experience in offline datasets can play a large role in the final performance of offline RL algorithms. Therefore in OG-MARL, for each environment and scenario, we include a range of dataset distributions including Good, Medium, Poor and Replay datasets in order to benchmark offline MARL algorithms on a range of different dataset qualities. For more information on why we chose to include each environment and its task properties, please read our accompanying paper.

Our datasets are now hosted on Hugging Face for improved accessibility for the community: https://huggingface.co/datasets/InstaDeepAI/og-marl

⚠️ Some datasets have yet to be converted to the new dataset format (Vault). For available datasets, please refer to og_marl/vault_utils/download_vault.py or the Hugging Face datasets repository.

Environments and Scenarios in OG-MARL 🗺️

Environment	Scenario	Agents	Act	Obs	Reward	Types	Repo
🔫SMAC v1	3m 8m 2s3z 5m_vs_6m 27m_vs_30m 3s5z_vs_3s6z 2c_vs_64zg	3 8 5 5 27 8 2	Discrete	Vector	Dense	Homog Homog Heterog Homog Homog Heterog Homog	source
💣SMAC v2	terran_5_vs_5 zerg_5_vs_5 terran_10_vs_10	5 5 10	Discrete	Vector	Dense	Heterog	source
🚅Flatland	3 Trains 5 Trains	3 5	Discrete	Vector	Sparse	Homog	source
🐜MAMuJoCo	2x3 HalfCheetah 2x4 Ant 4x2 Ant	2 2 4	Cont.	Vector	Dense	Heterog Homog Homog	source
🐻PettingZoo	Pursuit Co-op Pong	8 2	Discrete Discrete	Pixels Pixels	Dense	Homog Heterog	source

Datasets from Prior Works 🥇

We recently converted several datasets from prior works to Vaults and benchmarked our baseline algorithms on them. For more information, see our technical report on ArXiv.

Paper	Environment	Scenario	Source
Pan et al. (2022)	🐜MAMuJoCo	2x3 HalfCheetah	source
Pan et al. (2022)	🔴MPE	simple_spread	source
Shao et al. (2023)	🔫SMAC v1	5m_vs_6m 2s3z 3s_vs_5z 6h_vs_8z	source
Wang et al. (2023)	🔫SMAC v1	5m_vs_6m 6h_vs_8z 2c_vs_64zg corridor	source
Wang et al. (2023)	🐜MAMuJoCo	6x1 HalfCheetah 3x1 Hopper 2x4 Ant	source

Overview All Datasets

{"og_marl": {
        "smac_v1": {
            "3m": ["Good", "Medium", "Poor"],
            "8m": ["Good", "Medium", "Poor"],
            "5m_vs_6m": ["Good", "Medium", "Poor"],
            "2s3z": ["Good", "Medium", "Poor"],
            "3s5z_vs_3s6z": ["Good", "Medium", "Poor"],
        },
        "smac_v2": {
            "terran_5_vs_5": ["Replay"],
            "terran_10_vs_10": ["Replay"],
            "zerg_5_vs_5": ["Replay"],
        },
        "mamujoco": {
            "2halfcheetah": ["Good", "Medium", "Poor"]
        },
        "gymnasium_mamujoco": {
            "2ant": ["Replay"],
            "2halfcheetah": ["Replay"],
            "2walker": ["Replay"],
            "3hopper": ["Replay"],
            "4ant": ["Replay"],
            "6halfcheetah": ["Replay"],
        },
    },
    "cfcql": {
        "smac_v1": {
            "6h_vs_8z": ["Expert", "Medium", "Medium-Replay", "Mixed"],
            "3s_vs_5z": ["Expert", "Medium", "Medium-Replay", "Mixed"]
            "5m_vs_6m": ["Expert", "Medium", "Medium-Replay", "Mixed"]
            "2s3z": ["Expert", "Medium", "Medium-Replay", "Mixed"]
        },
    },
    "alberdice": {
        "rware": {
            "small-2ag": ["Expert"],
            "small-4ag": ["Expert"],
            "small-6ag": ["Expert"],
            "tiny-2ag": ["Expert"],
            "tiny-4ag": ["Expert"],
            "tiny-6ag": ["Expert"],
        },
    },
    "omar": {
        "mpe": {
            "simple_spread": ["Expert", "Medium", "Medium-Replay", "Random"]
            "simple_tag": ["Expert", "Medium", "Medium-Replay", "Random"]
            "simple_world": ["Expert", "Medium", "Medium-Replay", "Random"]
        },
        "mamujoco": {
            "2halfcheetah": ["Expert", "Medium", "Medium-Replay", "Random"]
        },
    },
    "omiga": {
        "smac_v1": {
            "2c_vs_64zg": ["Good", "Medium", "Poor"],
            "6h_vs_8z": ["Good", "Medium", "Poor"],
            "5m_vs_6m": ["Good", "Medium", "Poor"],
            "corridor": ["Good", "Medium", "Poor"],
        },
        "mamujoco": {
            "6halfcheetah": ["Expert", "Medium", "Medium-Expert", "Medium-Replay"],
            "2ant": ["Expert", "Medium", "Medium-Expert", "Medium-Replay"],
            "3hopper": ["Expert", "Medium", "Medium-Expert", "Medium-Replay"],
        },
    },
}

Installing MAMuJoCo 🐆

The OG-MARL datasets use the latest version of MuJoCo (210). While the OMIGA and OMAR datasets use an older version (200). They each have different instalation instructions and should be installed in seperate virtual environments.

MAMuJoCo 210

bash install_environments/mujoco210.sh

pip install -r install_environments/requirements/mujoco.txt

pip install -r install_environments/requirements/mamujoco210.txt

MAMuJoCo 200

bash install_environments/mujoco200.sh

pip install -r install_environments/requirements/mujoco.txt

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:~/.mujoco/mujoco200/bin

pip install -r install_environments/requirements/mamujoco200.txt

See Also 🔎

InstaDeep's MARL ecosystem in JAX. In particular, we suggest users check out the following sister repositories:

• 🦁 Mava: a research-friendly codebase for distributed MARL in JAX.
• 🌴 Jumanji: a diverse suite of scalable reinforcement learning environments in JAX.
• 😎 Matrax: a collection of matrix games in JAX.
• 🔦 Flashbax: accelerated replay buffers in JAX.
• 📈 MARL-eval: standardised experiment data aggregation and visualisation for MARL.

Related. Other libraries related to accelerated MARL in JAX.

• 🦊 JaxMARL: accelerated MARL environments with baselines in JAX.
• ♟️ Pgx: JAX implementations of classic board games, such as Chess, Go and Shogi.
• 🔼 Minimax: JAX implementations of autocurricula baselines for RL.

Citing OG-MARL

If you use OG-MARL Datasets in your work, please cite the library using:

@inproceedings{formanek2023ogmarl,
    author = {Formanek, Claude and Jeewa, Asad and Shock, Jonathan and Pretorius, Arnu},
    title = {Off-the-Grid MARL: Datasets and Baselines for Offline Multi-Agent Reinforcement Learning},
    year = {2023},
    publisher = {AAMAS},
    booktitle = {Extended Abstract at the 2023 International Conference on Autonomous Agents and Multiagent Systems},
}

Acknowledgements 🙏

The development of this library was supported with Cloud TPUs from Google's TPU Research Cloud (TRC) 🌤.