MARLlib: A benchmark for multi-agent reinforcement learning

Multi-Agent RLlib (MARLlib) is a MARL benchmark based on Ray and one of its toolkits RLlib. It provides MARL research community a unified platform for developing and evaluating the new ideas in various multi-agent environments. There are four core features of MARLlib.

• it collects most of the existing MARL algorithms widely acknowledged by the community and unifies them under one framework.
• it gives a solution that enables different multi-agent environments using the same interface to interact with the agents.
• it guarantees excellent efficiency in both the training and sampling process.
• it provides trained results, including learning curves and pretrained models specific to each task and algorithm's combination, with finetuned hyper-parameters to guarantee credibility.

Project Website: https://sites.google.com/view/marllib

Overview

We collected most of the existing multi-agent environment and multi-agent reinforcement learning algorithms and unified them under one framework based on Ray 's RLlib to boost the MARL research.

The MARL baselines include independence learning (IQL, A2C, DDPG, TRPO, PPO), centralized critic learning (COMA, MADDPG, MAPPO, HATRPO), and value decomposition (QMIX, VDN, FACMAC, VDA2C) are all implemented.

Popular environments like SMAC, MaMujoco, and Google Research Football are provided with a unified interface.

The algorithm code and environment code are fully separated. Changing the environment needs no modification on the algorithm side and vice versa.

Here we provide a table for comparison of MARLlib and before benchmarks.

Benchmark	Github Stars	Learning Mode	Available Env	Algorithm Type	Algorithm Number	Continues Control	Asynchronous Interact	Distributed Training	Framework	Last Update
PyMARL		CP	1	VD	5				*
PyMARL2		CP	1	VD	12				PyMARL
MARL-Algorithms		CP	1	VD+Comm	9				*
EPyMARL		CP	4	IL+VD+CC	10				PyMARL
Marlbenchmark		CP+CL	4	VD+CC	5	✔️			pytorch-a2c-ppo-acktr-gail
MAlib		SP	8	SP	9	✔️			*
MARLlib		CP+CL+CM+MI	10	IL+VD+CC	18	✔️	✔️	✔️	Ray/RLlib

CP, CL, CM, and MI represent cooperative, collaborative, competitive, and mixed task learning modes. IL, VD, and CC represent independent learning, value decomposition, and centralized critic categorization. SP represents self-play. Comm represents communication-based learning. Asterisk denotes that the benchmark uses its framework.

The tutorial of RLlib can be found at this link. Fast examples can be found at this link. These will help you quickly dive into RLlib.

We hope MARLlib can benefit everyone interested in RL/MARL.

Environment

Supported Multi-agent Environments / Tasks

Most of the popular environment in MARL research has been incorporated in this benchmark:

Env Name	Learning Mode	Observability	Action Space	Observations
LBF	Mixed	Both	Discrete	Discrete
RWARE	Collaborative	Partial	Discrete	Discrete
MPE	Mixed	Both	Both	Continuous
SMAC	Cooperative	Partial	Discrete	Continuous
MetaDrive	Collaborative	Partial	Continuous	Continuous
MAgent	Mixed	Partial	Discrete	Discrete
Pommerman	Mixed	Both	Discrete	Discrete
MaMujoco	Cooperative	Partial	Continuous	Continuous
GRF	Collaborative	Full	Discrete	Continuous
Hanabi	Cooperative	Partial	Discrete	Discrete

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

Algorithm

We provide three types of MARL algorithms as our baselines including:

Independent Learning: IQL DDPG PG A2C TRPO PPO

Centralized Critic: COMA MADDPG MAAC MAPPO MATRPO HATRPO HAPPO

Value Decomposition: VDN QMIX FACMAC VDAC VDPPO

Here is a chart describing the characteristics of each algorithm:

Algorithm	Support Task Mode	Need Global State	Action	Learning Mode	Type
IQL*	Mixed	No	Discrete	Independent Learning	Off Policy
PG	Mixed	No	Both	Independent Learning	On Policy
A2C	Mixed	No	Both	Independent Learning	On Policy
DDPG	Mixed	No	Continuous	Independent Learning	Off Policy
TRPO	Mixed	No	Both	Independent Learning	On Policy
PPO	Mixed	No	Both	Independent Learning	On Policy
COMA	Mixed	Yes	Both	Centralized Critic	On Policy
MADDPG	Mixed	Yes	Continuous	Centralized Critic	Off Policy
MAA2C*	Mixed	Yes	Both	Centralized Critic	On Policy
MATRPO*	Mixed	Yes	Both	Centralized Critic	On Policy
MAPPO	Mixed	Yes	Both	Centralized Critic	On Policy
HATRPO	Cooperative	Yes	Both	Centralized Critic	On Policy
HAPPO	Cooperative	Yes	Both	Centralized Critic	On Policy
VDN	Cooperative	No	Discrete	Value Decomposition	Off Policy
QMIX	Cooperative	Yes	Discrete	Value Decomposition	Off Policy
FACMAC	Cooperative	Yes	Continuous	Value Decomposition	Off Policy
VDAC	Cooperative	Yes	Both	Value Decomposition	On Policy
VDPPO*	Cooperative	Yes	Both	Value Decomposition	On Policy

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.

Current Task & Available algorithm mapping: Y for available, N for not suitable, P for partially available on some scenarios. (Note: in our code, independent algorithms may not have I as prefix. For instance, PPO = IPPO)

Env w Algorithm	IQL	PG	A2C	DDPG	TRPO	PPO	COMA	MADDPG	MAAC	MATRPO	MAPPO	HATRPO	HAPPO	VDN	QMIX	FACMAC	VDAC	VDPPO
LBF	Y	Y	Y	N	Y	Y	Y	N	Y	Y	Y	Y	Y	P	P	P	P	P
RWARE	Y	Y	Y	N	Y	Y	Y	N	Y	Y	Y	Y	Y	Y	Y	Y	Y	Y
MPE	P	Y	Y	P	Y	Y	P	P	Y	Y	Y	Y	Y	Y	Y	Y	Y	Y
SMAC	Y	Y	Y	N	Y	Y	Y	N	Y	Y	Y	Y	Y	Y	Y	Y	Y	Y
MetaDrive	N	Y	Y	Y	Y	Y	N	N	N	N	N	N	N	N	N	N	N	N
MAgent	Y	Y	Y	N	Y	Y	Y	N	Y	Y	Y	Y	Y	N	N	N	N	N
Pommerman	Y	Y	Y	N	Y	Y	P	N	Y	Y	Y	Y	Y	P	P	P	P	P
MaMujoco	N	Y	Y	Y	Y	Y	N	Y	Y	Y	Y	Y	Y	N	N	Y	Y	Y
GRF	Y	Y	Y	N	Y	Y	Y	N	Y	Y	Y	Y	Y	Y	Y	Y	Y	Y
Hanabi	Y	Y	Y	N	Y	Y	Y	N	Y	Y	Y	Y	Y	N	N	N	N	N

You can find a comprehensive list of existing MARL algorithms in different environments here.

Installation

Install Ray

pip install ray==1.8.0 # version sensitive

Add patch of MARLlib

cd patch
python add_patch.py

Y to replace source-packages code

Attention: Above is the common installation. Each environment needs extra dependency. Please read the installation instruction in envs/base_env/install.

Examples

python marl/main.py --algo_config=MAPPO [--finetuned] --env-config=smac with env_args.map_name=3m

--finetuned is optional, force using the finetuned hyperparameter

Navigation

We provide an introduction to the code directory to help you get familiar with the codebase:

• top level directory structure

This picture is in image/code-MARLlib.png

• MARL directory structure

This picture is in image/code-MARL.png.png

• ENVS directory structure

This picture is in image/code-ENVS.png.png

Contribute

MARLlib is friendly to incorporating a new environment. Besides the ten we already implemented, we support almost all kinds of MARL environments. Before contributing new environment, you must know:

Things you ought to cover:

• provide a new environment interface python file, follow the style of MARLlib/envs/base_env
• provide a corresponding config yaml file, follow the style of MARLlib/envs/base_env/config
• provide a corresponding instruction readme file, follow the style of MARLlib/envs/base_env/install

Things not essential:

• change the MARLlib data processing pipeline
• provide a unique runner or controller specific to the environment
• concern about the data logging

The ten environments we already contained have covered great diversity in action space, observation space, agent-env interaction style, task mode, additional information like action mask, etc. The best practice to incorporate your environment is to find an existing similar one and provide the same interface.

Bug Shooting

Our patch files fix most RLlib-related errors on MARL.

Here we only list the common bugs, not RLlib-related. (Mostly is your mistake)

• observation/action out of space bug:

  • make sure the observation/action space defined in env init function
    • has same data type with env returned data (e.g., float32/64)
    • env returned data range is in the space scope (e.g., box(-2,2))
  • the returned env observation contained the required key (e.g., action_mask/state)

• Action NaN is invaild bug

  • this is common bug espectially in continues control problem, carefully finetune the algorithm's hyperparameter
    • smaller learning rate
    • set some action value bound