MARLlib: A Multi-agent Reinforcement Learning Library

 

❗ News
March 2023 ⚓We are excited to announce that a major update has just been released. For detailed version information, please refer to the version info.
May 2023 Exciting news! MARLlib now supports five more tasks: MATE, GoBigger, Overcooked-AI, MAPDN, and AirCombat. Give them a try!
June 2023 OpenAI: Hide and Seek and SISL environments are incorporated into MARLlib.
Aug 2023 🎉MARLlib has been accepted for publication in JMLR.
Sept 2023 Latest PettingZoo with Gymnasium are compatiable within MARLlib.
Nov 2023 We are currently in the process of creating a hands-on MARL book and aim to release the draft by the end of 2023.

Multi-agent Reinforcement Learning Library (MARLlib) is a MARL library that utilizes Ray and one of its toolkits RLlib. It offers a comprehensive platform for developing, training, and testing MARL algorithms across various tasks and environments.

Here's an example of how MARLlib can be used:

from marllib import marl

# prepare env
env = marl.make_env(environment_name="mpe", map_name="simple_spread", force_coop=True)

# initialize algorithm with appointed hyper-parameters
mappo = marl.algos.mappo(hyperparam_source='mpe')

# build agent model based on env + algorithms + user preference
model = marl.build_model(env, mappo, {"core_arch": "mlp", "encode_layer": "128-256"})

# start training
mappo.fit(env, model, stop={'timesteps_total': 1000000}, share_policy='group')

Why MARLlib?

Here we provide a table for the comparison of MARLlib and existing work.

Library	Supported Env	Algorithm	Parameter Sharing	Model
PyMARL	1 cooperative	5	share	GRU
PyMARL2	2 cooperative	11	share	MLP + GRU
MAPPO Benchmark	4 cooperative	1	share + separate	MLP + GRU
MAlib	4 self-play	10	share + group + separate	MLP + LSTM
EPyMARL	4 cooperative	9	share + separate	GRU
HARL	8 cooperative	9	share + separate	MLP + CNN + GRU
MARLlib	17 no task mode restriction	18	share + group + separate + customizable	MLP + CNN + GRU + LSTM

Library	Github Stars	Documentation	Issues Open	Activity	Last Update
PyMARL		❌
PyMARL2		❌
MAPPO Benchmark		❌
MAlib
EPyMARL		❌
HARL*		❌
MARLlib

* HARL is the latest MARL library that has been recently released:fire:. If cutting-edge MARL algorithms with state-of-the-art performance are your target, HARL is definitely worth a look!

key features

🔰 MARLlib offers several key features that make it stand out:

• MARLlib unifies diverse algorithm pipelines with agent-level distributed dataflow, allowing researchers to develop, test, and evaluate MARL algorithms across different tasks and environments.
• MARLlib supports all task modes, including cooperative, collaborative, competitive, and mixed. This makes it easier for researchers to train and evaluate MARL algorithms across a wide range of tasks.
• MARLlib provides a new interface that follows the structure of Gym, making it easier for researchers to work with multi-agent environments.
• MARLlib provides flexible and customizable parameter-sharing strategies, allowing researchers to optimize their algorithms for different tasks and environments.

🚀 Using MARLlib, you can take advantage of various benefits, such as:

• Zero knowledge of MARL: MARLlib provides 18 pre-built algorithms with an intuitive API, allowing researchers to start experimenting with MARL without prior knowledge of the field.
• Support for all task modes: MARLlib supports almost all multi-agent environments, making it easier for researchers to experiment with different task modes.
• Customizable model architecture: Researchers can choose their preferred model architecture from the model zoo, or build their own.
• Customizable policy sharing: MARLlib provides grouping options for policy sharing, or researchers can create their own.
• Access to over a thousand released experiments: Researchers can access over a thousand released experiments to see how other researchers have used MARLlib.

Installation

Note: Please note that at this time, MARLlib is only compatible with Linux operating systems.

Step-by-step (recommended)

• install dependencies
• install environments
• install patches

1. install dependencies (basic)

First, install MARLlib dependencies to guarantee basic usage. following this guide, finally install patches for RLlib.

$ conda create -n marllib python=3.8 # or 3.9
$ conda activate marllib
$ git clone https://github.com/Replicable-MARL/MARLlib.git && cd MARLlib
$ pip install -r requirements.txt

2. install environments (optional)

Please follow this guide.

Note: We recommend the gym version around 0.20.0.

pip install "gym==0.20.0"

3. install patches (basic)

Fix bugs of RLlib using patches by running the following command:

$ cd /Path/To/MARLlib/marllib/patch
$ python add_patch.py -y

PyPI

$ pip install --upgrade pip
$ pip install marllib

Docker-based usage

We provide a Dockerfile for building the MARLlib docker image in MARLlib/docker/Dockerfile and a devcontainer setup in MARLlib/.devcontainer folder. If you use the devcontainer, one thing to note is that you may need to customise certain arguments in runArgs of devcontainer.json according to your hardware, for example the --shm-size argument.

Getting started

Prepare the configuration

There are four parts of configurations that take charge of the whole training process.

• scenario: specify the environment/task settings
• algorithm: choose the hyperparameters of the algorithm
• model: customize the model architecture
• ray/rllib: change the basic training settings

Before training, ensure all the parameters are set correctly, especially those you don't want to change.

Note: You can also modify all the pre-set parameters via MARLLib API.*

Register the environment

Ensure all the dependencies are installed for the environment you are running with. Otherwise, please refer to MARLlib documentation.

task mode	api example
cooperative	marl.make_env(environment_name="mpe", map_name="simple_spread", force_coop=True)
collaborative	marl.make_env(environment_name="mpe", map_name="simple_spread")
competitive	marl.make_env(environment_name="mpe", map_name="simple_adversary")
mixed	marl.make_env(environment_name="mpe", map_name="simple_crypto")

Most of the popular environments in MARL research are supported by MARLlib:

Env Name	Learning Mode	Observability	Action Space	Observations
LBF	cooperative + collaborative	Both	Discrete	1D
RWARE	cooperative	Partial	Discrete	1D
MPE	cooperative + collaborative + mixed	Both	Both	1D
SISL	cooperative + collaborative	Full	Both	1D
SMAC	cooperative	Partial	Discrete	1D
MetaDrive	collaborative	Partial	Continuous	1D
MAgent	collaborative + mixed	Partial	Discrete	2D
Pommerman	collaborative + competitive + mixed	Both	Discrete	2D
MAMuJoCo	cooperative	Full	Continuous	1D
GRF	collaborative + mixed	Full	Discrete	2D
Hanabi	cooperative	Partial	Discrete	1D
MATE	cooperative + mixed	Partial	Both	1D
GoBigger	cooperative + mixed	Both	Continuous	1D
Overcooked-AI	cooperative	Full	Discrete	1D
PDN	cooperative	Partial	Continuous	1D
AirCombat	cooperative + mixed	Partial	MultiDiscrete	1D
HideAndSeek	competitive + mixed	Partial	MultiDiscrete	1D

Each environment has a readme file, standing as the instruction for this task, including env settings, installation, and important notes.

Initialize the algorithm

running target	api example
train & finetune	marl.algos.mappo(hyperparam_source=$ENV)
develop & debug	marl.algos.mappo(hyperparam_source="test")
3rd party env	marl.algos.mappo(hyperparam_source="common")

Here is a chart describing the characteristics of each algorithm:

algorithm	support task mode	discrete action	continuous action	policy type
IQL*	all four	✔️		off-policy
PG	all four	✔️	✔️	on-policy
A2C	all four	✔️	✔️	on-policy
DDPG	all four		✔️	off-policy
TRPO	all four	✔️	✔️	on-policy
PPO	all four	✔️	✔️	on-policy
COMA	all four	✔️		on-policy
MADDPG	all four		✔️	off-policy
MAA2C*	all four	✔️	✔️	on-policy
MATRPO*	all four	✔️	✔️	on-policy
MAPPO	all four	✔️	✔️	on-policy
HATRPO	cooperative	✔️	✔️	on-policy
HAPPO	cooperative	✔️	✔️	on-policy
VDN	cooperative	✔️		off-policy
QMIX	cooperative	✔️		off-policy
FACMAC	cooperative		✔️	off-policy
VDAC	cooperative	✔️	✔️	on-policy
VDPPO*	cooperative	✔️	✔️	on-policy

*all four: cooperative collaborative competitive mixed

IQL is the multi-agent version of Q learning. MAA2C and MATRPO are the centralized version of A2C and TRPO. VDPPO is the value decomposition version of PPO.

Build the agent model

An agent model consists of two parts, encoder and core arch. encoder will be constructed by MARLlib according to the observation space. Choose mlp, gru, or lstm as you like to build the complete model.

model arch	api example
MLP	marl.build_model(env, algo, {"core_arch": "mlp")
GRU	marl.build_model(env, algo, {"core_arch": "gru"})
LSTM	marl.build_model(env, algo, {"core_arch": "lstm"})
Encoder Arch	marl.build_model(env, algo, {"core_arch": "gru", "encode_layer": "128-256"})

Kick off the training

setting	api example
train	algo.fit(env, model)
debug	algo.fit(env, model, local_mode=True)
stop condition	algo.fit(env, model, stop={'episode_reward_mean': 2000, 'timesteps_total': 10000000})
policy sharing	algo.fit(env, model, share_policy='all') # or 'group' / 'individual'
save model	algo.fit(env, model, checkpoint_freq=100, checkpoint_end=True)
GPU accelerate	algo.fit(env, model, local_mode=False, num_gpus=1)
CPU accelerate	algo.fit(env, model, local_mode=False, num_workers=5)

Training & rendering API

from marllib import marl

# prepare env
env = marl.make_env(environment_name="smac", map_name="5m_vs_6m")
# initialize algorithm with appointed hyper-parameters
mappo = marl.algos.mappo(hyperparam_source="smac")
# build agent model based on env + algorithms + user preference
model = marl.build_model(env, mappo, {"core_arch": "gru", "encode_layer": "128-256"})
# start training
mappo.fit(
  env, model, 
  stop={"timesteps_total": 1000000}, 
  checkpoint_freq=100, 
  share_policy="group"
)
# rendering
mappo.render(
  env, model, 
  local_mode=True, 
  restore_path={'params_path': "checkpoint/params.json",
                'model_path': "checkpoint/checkpoint-10"}
)

Results

Under the current working directory, you can find all the training data (logging and TensorFlow files) as well as the saved models. To visualize the learning curve, you can use Tensorboard. Follow the steps below:

1. Install Tensorboard by running the following command:

pip install tensorboard

2. Use the following command to launch Tensorboard and visualize the results:

tensorboard --logdir .

Alternatively, you can refer to this tutorial for more detailed instructions.

For a list of all the existing results, you can visit this link. Please note that these results were obtained from an older version of MARLlib, which may lead to inconsistencies when compared to the current results.

Quick examples

MARLlib provides some practical examples for you to refer to.

• Detailed API usage: show how to use MARLlib api in detail, e.g. cmd + api combined running.
• Policy sharing cutomization: define your group policy-sharing strategy as you like based on current tasks.
• Loading model: load the pre-trained model and keep training.
• Loading model and rendering: render the environment based on the pre-trained model.
• Incorporating new environment: add your new environment following MARLlib's env-agent interaction interface.
• Incorporating new algorithm: add your new algorithm following MARLlib learning pipeline.
• Parallelized finetuning: fintune your policy/model performance with ray.tune.

Tutorials

Try MPE + MAPPO examples on Google Colaboratory! More tutorial documentations are available here.

Awesome List

A collection of research and review papers of multi-agent reinforcement learning (MARL) is available. The papers have been organized based on their publication date and their evaluation of the corresponding environments.

Algorithms: Environments:

Community

Channel	Link
Issues	GitHub Issues

Roadmap

The roadmap to the future release is available in ROADMAP.md.

Contributing

We are a small team on multi-agent reinforcement learning, and we will take all the help we can get! If you would like to get involved, here is information on contribution guidelines and how to test the code locally.

You can contribute in multiple ways, e.g., reporting bugs, writing or translating documentation, reviewing or refactoring code, requesting or implementing new features, etc.

Citation

If you use MARLlib in your research, please cite the MARLlib paper.

@article{hu2022marllib,
  author  = {Siyi Hu and Yifan Zhong and Minquan Gao and Weixun Wang and Hao Dong and Xiaodan Liang and Zhihui Li and Xiaojun Chang and Yaodong Yang},
  title   = {MARLlib: A Scalable and Efficient Multi-agent Reinforcement Learning Library},
  journal = {Journal of Machine Learning Research},
  year    = {2023},
}

Works that are based on or closely collaborate with MARLlib <link>

@InProceedings{hu2022policy,
      title={Policy Diagnosis via Measuring Role Diversity in Cooperative Multi-agent {RL}},
      author={Hu, Siyi and Xie, Chuanlong and Liang, Xiaodan and Chang, Xiaojun},
      booktitle={Proceedings of the 39th International Conference on Machine Learning},
      year={2022},
}
@misc{zhong2023heterogeneousagent,
      title={Heterogeneous-Agent Reinforcement Learning}, 
      author={Yifan Zhong and Jakub Grudzien Kuba and Siyi Hu and Jiaming Ji and Yaodong Yang},
      archivePrefix={arXiv},
      year={2023},
}