README.md

Aggresiveness-regulated Multi-agent Stress Testing of Autonomous Vehicles

This repo dedicates to the task of stress testing a black-box autnomous vehicle system using multi-agent reinforcement learnign algorithms under an adversarial environment.

The simulation environment used are included as a submodule, which is adapted from highway-env.

The codes in this repo follow the suggested one-file implementation in reinforcement learning community.

There are four algorithms experimented in this repo:

• MACPO (Multi-Agent Constrained Policy Optimization)
• HATRPO (Heterogeneous-Agent Trust Region Policy Optimization)
• CPO (Constrained Policy Optimization)
• TRPO (Trust Region Policy Optimization)

The results are summarized in this thesis

Environment Setup

The python environemnt can be created using conda:

conda env create -f environment.yml

Running

To run the training script of an algorithm ALGO listed above ALGO.py:

python ALGO.py

To run the testing script of an algorithm ALGO listed above ALGO_vis.py:

python ALGO_vis.py

Acknowledgments

The code within this repository was developed with assistance from MACPO and CleanRL.