PyTorch Implementation of COLA-MADDPG based on MADDPG-pytorch.
- OpenAI baselines, commit hash: 98257ef8c9bd23a24a330731ae54ed086d9ce4a7
- Multi-agent Particle Environments
- PyTorch
- OpenAI Gym, version: 0.9.4
- Tensorboard
All training code is contained within main.py
. To view options simply run:
python main.py --help
For vanilla MADDPG:
python main.py simple_tag_coop examplemodel --n_episodes 20000
For COLA-MADDPG:
python main.py simple_tag_coop examplemodel --n_episodes 20000 --consensus
-
To install,
cd
intomultiagent-particle-envs
directory and typepip install -e .
-
To interactively view moving to landmark scenario (see others in ./scenarios/):
bin/interactive.py --scenario simple.py
-
Known dependencies: OpenAI gym, numpy
-
To use the environments, look at the code for importing them in
make_env.py
.
The three scenarios we used in the paper are simple_tag_coop
, simple_spread
, and simple_reference_no_comm
. They correspond to "Cooperative Predator-Prey", "Cooperative Navigation" and "Cooperative Pantomime" in the text, respectively.
The OpenAI baselines Tensorflow implementation and Ilya Kostrikov's Pytorch implementation of DDPG were used as references. After the majority of this codebase was complete, OpenAI released their code for MADDPG, and I made some tweaks to this repo to reflect some of the details in their implementation (e.g. gradient norm clipping and policy regularization).