DeepCollision: Learning Configurations of Operating Environment of Autonomous Vehicles to Maximize their Collisions

With the aim to test autonomous driving systems, we propose a novel reinforcement learning (RL)-based approach named DeepCollision to learn operating environment configurations of autonomous vehicles, including formalizing environment configuration learning as an MDP and adopting DQN algorithm as the RL solution; DeepCollision learns environment configurations to maximize collisions of an Autonomous Vehicle Under Test (AVUT).

This dataset contains:

1. algorithms - The algorithm of DeepCollision, which includes the network architecture and the DQN hyperparameter settings;
2. pilot-study - All the raw data and plots for the pilot study;
3. formal-experiment - A dataset contains all the raw data for analysis and the scenarios with detailed demand values;
4. rest-api - The REST API endpoints for environment configuration and one example to show the usage of the APIs.