AutoEncodingBayesianInverseGames.jl

This repository contains code for our paper at WAFR 2024: Auto-Encoding Bayesian Inverse Games. We propose an end-to-end differentiable pipeline that embeds a differentiable Nash game solver into a generative model for multi-modal uncertainty inference in multi-agent scenarios. For more information, please visit our project website.

To differentiate through a game solver, we analytically derive gradient from the optimality conditions of the game based on the implicit function theorem. For more details, please refer to our prior work.

How to use

Initial setup

• Install Julia 1.10.5

• Set up a license key for the PATHSolver (see License section below)

• Start Julia in the root directory by typing julia in the terminal

• Activate the package environment by hitting ] to enter the package mode first and then type: activate .

• Instantiate the environment in the package mode if you haven't done so before by typing instantiate

• Exit the package mode by hitting the backspace key; precompile the package: using DrivingExample

Run simulation

Run traffic intersection simulation with a pre-trained VAE model in the data\ folder:

DrivingExample.run_intersection_inference()

The generated videos will be stored in the data\ folder

Run model training

Generate a traffic dataset:

DrivingExample.run_intersection_data_collection()

Run VAE training:

DrivingExample.train_generative_model_with_driving_data()

The generated objects will be stored in the data\ folder

Cite this work

@article{liu2024auto,
  title={Auto-Encoding Bayesian Inverse Games},
  author={Liu, Xinjie and Peters, Lasse and Alonso-Mora, Javier and Topcu, Ufuk and Fridovich-Keil, David},
  journal={arXiv preprint arXiv:2402.08902},
  year={2024}
}

Acknowledgements

This project adapts the following packages as libraries:

• Contingency Games

• DifferentiableAdaptiveGames.jl

License

This package uses PATH solver (via PATHSolver.jl) under the hood. Larger-sized problems require to have a license key. By courtesy of Steven Dirkse, Michael Ferris, and Tudd Munson, temporary license keys are available free of charge. For more details about the license key, please consult PATHSolver.jl (License section).

Note that when no license is loaded, PATH does not report an informative error and instead may just report a wrong result. Thus, please make sure that the license is loaded correctly before using the solver.

---

TODO: Multi-processing part for accelerating large-scale training