Inverse Mixed Strategy Games with Generative Trajectory Models

Authors: Max Muchen Sun, Pete Trautman, Todd Murphey

This repository contains the implementation for the paper "Inverse Mixed Strategy Games with Generative Trajectory Models" (PDF), which will be presented at 2025 International Conference on Robotics and Automation (ICRA).

In this work, we propose an inverse game method that integrates a generative trajectory model into a differentiable mixed-strategy game framework--Bayesian Recursive Nash Equilibrium (BRNE). By representing the mixed strategy with a conditional variational autoencoder (CVAE), our method can infer high-dimensional, multi-modal behavior distributions from noisy measurements while adapting in real-time to new observations. More information (including tutorials) on BRNE can be found at this indepedent repository.

This repository contains the following:

• dataset_generation.py: The script to generate synthetic multi-agent dataset using iLQGames, which uses the utitlity functions in dynax.py. We also include the pre-generated dataset in synthetic_dataset_5agents.zip.
• inverse_mixed_strategy.ipynb: This notebook contains all the steps for training the CVAE model and solving the inverse game with BRNE. We also provide the pre-trained weights for the CVAE (cvae_5agents(./dynax.py).pkl) and the MLP-based risk/interaction function (risk_mlp_5agents.pkl).

Contact: Please contact Max Muchen Sun (msun@u.northwestern.edu) if you have any question.