Multi-Agent POsthumous Credit Assignment (MA-POCA) is a multi-agent learning algorithm that can gracefully deal with spawning and despawning agents without the use of absorbing states.
You can check the original paper here.
This notebook can be used to generate some of the plots of the paper (Figure 1 and Appendix A).
To install the code for training, in the ma-poca folder, run the command:
pip install -e mapoca/
The mapoca package functions similarly to the mlagents package (see here for more information)
To launch an experiment, use the command :
mapoca-learn config/<algorithm>/<env_name>.yaml --env=<env_name> --run-id=<test-id>
where:
<env_name>is the name of one of our four environments:- BatonPass
- DungeonEscape
- ParticlesEnv
- PushBlockCollab
<algorithm>can be one of the three provided algorithms:- mapoca
- coma
- ppo
- and
<test-id>a unique identifier for the experiment
The results of the experiment will be inside the results folder.