GrAMMI: Graph Based Adversarial Modeling with Mutual Information

Arxiv - IROS 2023

GrAMMI is an adversarial tracking method that uses a graph neural network along with a regularized gaussian mixture model that is regularized using mutual information. The method is designed for tracking adversarial targets in sparse partially observable environments.

Training Models

1. Create Conda environment

conda env create -f requirements.yaml
conda activate grammi

2. Download Datasets

Visit the Google Drive to download the datasets. We provide datasets for the Smuggler Domain (high and low visibility) and Prisoner datasets (high, medium, and low visibility). Download the datasets to grammi_datasets in the same directory for the default options.

3. Train

The models can be trained using run_multiple.py. The following command trains multiple seeds for a single model type and a single time horizon prediction.

4. Evaluate the model

Once the models are trained, you can evaluate the models using evaluate/evaluate_models_mi.py. This will record the metrics for the models based on the test set.

Details of configuration

Citation

If you find our code or paper is useful, please consider citing:

@inproceedings{ye2023grammi,
  title={Learning Models of Adversarial Agent Behavior under Partial
Observability},
  author={Ye, Sean and Natarajan, Manisha and Wu, Zixuan and Paleja, Rohan and Chen, Letian and Gombolay, Matthew},
  booktitle={IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year={2023}
}

License

This code is distributed under an MIT LICENSE.