Extended Python MARL framework - EPyMARL

EPyMARL is an extension of PyMARL, and includes

Additional algorithms (IA2C, IPPO, MADDPG, MAA2C and MAPPO)
Support for Gym environments (on top of the existing SMAC support)
Option for no-parameter sharing between agents (original PyMARL only allowed for parameter sharing)
Flexibility with extra implementation details (e.g. hard/soft updates, reward standarization, and more)
Consistency of implementations between different algorithms (fair comparisons)

Extended Python MARL framework - EPyMARL
Table of Contents
Installation & Run instructions
- Installing LBF, RWARE, and MPE
- Using A Custom Gym Environment
Run an experiment on a Gym environment
Run a hyperparameter search
Saving and loading learnt models
- Saving models
- Loading models
Citing PyMARL and EPyMARL
License

Installation & Run instructions

For information on installing and using this codebase with SMAC, we suggest visiting and reading the original PyMARL README. Here, we maintain information on using the extra features EPyMARL offers. To install the codebase, clone this repo and install the requirements.txt.

Installing LBF, RWARE, and MPE

In Benchmarking Multi-Agent Deep Reinforcement Learning Algorithms in Cooperative Tasks we introduce and benchmark algorithms in Level-Based Foraging, Multi-Robot Warehouse and Multi-agent Particle environments. To install these please visit:

Level Based Foraging or install with pip install lbforaging
Multi-Robot Warehouse or install with pip install rware
Our fork of MPE, clone it and install it with pip install -e .

Example of using LBF:

python3 src/main.py --config=qmix --env-config=gymma with env_args.time_limit=25 env_args.key="lbforaging:Foraging-8x8-2p-3f-v1"

Example of using RWARE:

python3 src/main.py --config=qmix --env-config=gymma with env_args.time_limit=500 env_args.key="rware:rware-tiny-2ag-v1"

For MPE, our fork is needed. Essentially all it does (other than fixing some gym compatibility issues) is i) registering the environments with the gym interface when imported as a package and ii) correctly seeding the environments iii) makes the action space compatible with Gym (I think MPE originally does a weird one-hot encoding of the actions).

The environments names in MPE are:

...
    "multi_speaker_listener": "MultiSpeakerListener-v0",
    "simple_adversary": "SimpleAdversary-v0",
    "simple_crypto": "SimpleCrypto-v0",
    "simple_push": "SimplePush-v0",
    "simple_reference": "SimpleReference-v0",
    "simple_speaker_listener": "SimpleSpeakerListener-v0",
    "simple_spread": "SimpleSpread-v0",
    "simple_tag": "SimpleTag-v0",
    "simple_world_comm": "SimpleWorldComm-v0",
...

Therefore, after installing them you can run it using:

python3 src/main.py --config=qmix --env-config=gymma with env_args.time_limit=25 env_args.key="mpe:SimpleSpeakerListener-v0"

The pretrained agents are included in this repo here. You can use them with:

python3 src/main.py --config=qmix --env-config=gymma with env_args.time_limit=25 env_args.key="mpe:SimpleAdversary-v0" env_args.pretrained_wrapper="PretrainedAdversary"

and

python3 src/main.py --config=qmix --env-config=gymma with env_args.time_limit=25 env_args.key="mpe:SimpleTag-v0" env_args.pretrained_wrapper="PretrainedTag"

Using A Custom Gym Environment

EPyMARL supports environments that have been registered with Gym. The only difference with the Gym framework would be that the returned rewards should be a tuple (one reward for each agent). In this cooperative framework we sum these rewards together.

Environments that are supported out of the box are the ones that are registered in Gym automatically. Examples are: Level-Based Foraging and RWARE.

To register a custom environment with Gym, use the template below (taken from Level-Based Foraging).

from gym.envs.registration import registry, register, make, spec
register(
  id="Foraging-8x8-2p-3f-v1",                     # Environment ID.
  entry_point="lbforaging.foraging:ForagingEnv",  # The entry point for the environment class
  kwargs={
            ...                                   # Arguments that go to ForagingEnv's __init__ function.
        },
    )

Run an experiment on a Gym environment

python3 src/main.py --config=qmix --env-config=gymma with env_args.time_limit=50 env_args.key="lbforaging:Foraging-8x8-2p-3f-v1"

In the above command --env-config=gymma (in constrast to sc2 will use a Gym compatible wrapper). env_args.time_limit=50 sets the maximum episode length to 50 and env_args.key="..." provides the Gym's environment ID. In the ID, the lbforaging: part is the module name (i.e. import lbforaging will run automatically).

The config files act as defaults for an algorithm or environment.

They are all located in src/config. --config refers to the config files in src/config/algs --env-config refers to the config files in src/config/envs

All results will be stored in the Results folder.

Run a hyperparameter search

We include a script named search.py which reads a search configuration file (e.g. the included search.config.example.yaml) and runs a hyperparameter search in one or more tasks. The script can be run using

python search.py run --config=search.config.example.yaml --seeds 5 locally

In a cluster environment where one run should go to a single process, it can also be called in a batch script like:

python search.py run --config=search.config.example.yaml --seeds 5 single 1

where the 1 is an index to the particular hyperparameter configuration and can take values from 1 to the number of different combinations.

Saving and loading learnt models

Saving models

You can save the learnt models to disk by setting save_model = True, which is set to False by default. The frequency of saving models can be adjusted using save_model_interval configuration. Models will be saved in the result directory, under the folder called models. The directory corresponding each run will contain models saved throughout the experiment, each within a folder corresponding to the number of timesteps passed since starting the learning process.

Loading models

Learnt models can be loaded using the checkpoint_path parameter, after which the learning will proceed from the corresponding timestep.

Citing EPyMARL and PyMARL

The Extended PyMARL (EPyMARL) codebase was used in Benchmarking Multi-Agent Deep Reinforcement Learning Algorithms in Cooperative Tasks.

Georgios Papoudakis, Filippos Christianos, Lukas Schäfer, & Stefano V. Albrecht. Benchmarking Multi-Agent Deep Reinforcement Learning Algorithms in Cooperative Tasks, Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS), 2021

In BibTeX format:

@inproceedings{papoudakis2021benchmarking,
   title={Benchmarking Multi-Agent Deep Reinforcement Learning Algorithms in Cooperative Tasks},
   author={Georgios Papoudakis and Filippos Christianos and Lukas Schäfer and Stefano V. Albrecht},
   booktitle = {Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS)},
   year={2021},
   url = {http://arxiv.org/abs/2006.07869},
   openreview = {https://openreview.net/forum?id=cIrPX-Sn5n},
   code = {https://github.com/uoe-agents/epymarl},
}

If you use the original PyMARL in your research, please cite the SMAC paper.

M. Samvelyan, T. Rashid, C. Schroeder de Witt, G. Farquhar, N. Nardelli, T.G.J. Rudner, C.-M. Hung, P.H.S. Torr, J. Foerster, S. Whiteson. The StarCraft Multi-Agent Challenge, CoRR abs/1902.04043, 2019.