The repository consists of the code base for EMIX. EMIX is a multi-agent value factorization algorithm built on top of QMIX. The framework minimizes surprise by utilizing the energy across agents in the case of multi-agent partially-observable settings. By making use of a novel surprise value function in conjunction with an energy operator, EMIX minimizes the joint surprise across all agents in the multi-agent population.
The implementation makes use of PyMARL (writtent in PyTorch) and is built on top of the QMIX implementation.
[Skip this step if using a local machine]
Build the Dockerfile using
cd docker
bash build.sh
The following command will download SC2 in the 3rdparty folder and copy the custom SMAC maps required for experiments-
bash install_sc2.sh
Library dependencies are listed in the requirements.txt file and can be installed using the following command-
pip install -r requirements.txt
Once the installation is complete, EMIX experiments can be executed using the following command:
python3 prog/main.py --config=emix --env-config=sc2 with env_args.map_name=2s3z
The config files consist of default hyperparameter values. To change these for EMIX refer to emix.yaml config file located in prog/config folder.
To run experiments using the Docker container use the following-
bash run.sh $GPU python3 prog/main.py --config=emix --env-config=sc2 with env_args.map_name=2s3z
All results will be stored in the Results folder. For additional details on loading and saving models please refer to the PyMARL page.
Code licensed under the Apache License v2.0