Implementation of DMFG-QL and DMFG-AC in the AAAI-2022 paper - Decentralized Mean Field Games.

Our first set of environments pertain to simulated domains using the Petting Zoo simulator. The second set of experiments contains a real-world ride-sharing environment.

• See folder battle for training and testing scripts of the Battle environment.

• See folder combinedarms for training and testing scripts of the Combined Arms environment.

• See folder gather for training and testing scripts of the Gather environment.

• See folder tiger for training and testing scripts of the Tiger-Deer environment.

• See folder for waterworld for training and testing scripts of the Waterworld environment.

• See folder for ridesharing for training and testing scripts of the Ride sharing environment.

• /battle/pettingzoomagentDQNdecentralizedparallel.py: This file contains the training script for IL in the battle game. Similarly change the name of the algorithm to see the training of the 5 different algorithms.

• /battle/pettingzoomagentexecutionDMFGQLvsIL.py: This file contains the execution script for the battle between DMFG-QL and IL. Change the name of the algorithms to see the execution experiments between different algorithms.

Similarly the training and execution scripts for all the different environments can be found in the respective folders.

• /battle/RL_brainDQN.py: This file contains the IL implementation.

• /battle/RL_braindmfgql.py: This file contains the DMFG-QL implementation.

• /battle/RL_dmfgac.py: This file contains the DMFG-AC implementation.

• /battle/RL_brainMFQ.py: This file contains the MFQ implementation.

• /battle/RL_actorcritic.py: This file contains the MFAC implementation.

These algorithmic files are almost the same across the different environmental domains.

• /ridesharing/src2 - This folder contains the files for running DMFG-QL in the ride sharing environment. In this folder the file ValueFunction.py contains the algorithmic details.

• /ridesharing/src - This folder contains the files for running NeurADP and CO in the ride sharing environment. In this folder the file ValueFunction.py contains the algorithmic details.

In addition, we have a set of python files in the main directory pertaining to modifications of the environmental domains as compared to the defaults given in the Petting Zoo environments.

• python==3.7.1

sudo add-apt-repository ppa:deadsnakes/ppa
sudo apt-get update
sudo apt-get install python3.7

• Gym environment (0.18.0)

pip install gym==0.18.0

• MAgent environment

pip install magent

Replace the gridworld.py file in the magent installation with the file provided in this repository.

• Petting zoo

Please do not install petting zoo from source. Our implementation contains some changes from the souce implementation. The important files that have been changed are provided in the root directory for simplicity and easy reference.

cd pettingzoo
pip install ./ 

• tensorflow 2

pip install --upgrade pip
pip install tensorflow

• Pandas

pip install pandas 

• python==3.6.0

sudo add-apt-repository ppa:deadsnakes/ppa
sudo apt-get update
sudo apt-get install python3.6

• Tensorflow 1.12

pip install tensorflow==1.12

• Keras 2.2.4

pip install keras==2.2.4

• CPLEX 12.8 (Note: Commercial license is required): Follow the instructions here

• Training

cd battle
python pettingzoomagentDQNdecentralizedparallel.py 

This will run the training for the battle domain with the IL algorithm. Similarly run the other files for training other algorithms. For training other environments, change to the respective folders.

• Testing

python pettingzoomagentexecutionDMFGQLvsIL.py      

The above command is for running the test battles between the DFMG-QL and IL algorithms. Similarly run the respective execution script for the battles. Make sure all the training is conducted first and the training files are stored in a suitable folder. Include the saved file information while running the respective scripts.

• For running CO or NeurADP algorithms.

cd ridesharing/src
python main.py 

• For running DMFG-QL algorithm

cd ridesharing/src2
python main.py 

Due to copyright and size restrictions, the processed dataset used for the ride sharing experiments cannot be released here. Please email [email protected] for obtaining this data.

We use a set of 30 random seeds (1 -- 30) for all training experiments and a new set of 30 random seeds (31 -- 60) for all execution experiments.

This is research code and will not be actively maintained. Please send an email to [email protected] for questions or comments.

We would like to cite the MAgent for the MAgent simulator on which most of our virtual experiments are based. We are using the modified version of the games from the Pettingzoo environment. We would also like to cite Reinforcement Learning and Tensorflow from which we framed the structure of our algorithms and obtained the required baselines. Importantly, we would like to acknowledge Sanket Shah for data used in the ride sharing experiments. Our algorithmic implementations for the ride sharing experiments are based on the architectures and code base provided by Sanket Shah.

If you found this helpful, please cite the following paper:

@InProceedings{Sriramdmfg2022,
  title = 	 {Decentralized Mean Field Games},
  author = 	 {Subramanian, Sriram Ganapathi and Taylor, Matthew E. and Crowley, Mark and Poupart, Pascal} 
  booktitle = 	 {Proceedings of the Association for the Advancement of Artificial Intelligence (AAAI-2022)},
  year = 	 {2022},
  editor = 	 {Katia Sycara},
  address = 	 {Vancouver, Canada},
  month = 	 {Feb 22 -- 1 March},
  publisher = 	 {AAAI press}
}