Edge Case searching for Autonomous Vehicles

Sammanfattning

In this thesis, a method that generates critical situations for autonomous vehicles with fewer resources was investigated. To produce these key examples, the strategy in this thesis explores optimization and reinforcement learning. These key examples are defined as Edge cases; these cases are situations on the border between safe and unsafe. To be able to search for them, in this paper, they are defined numerically with Time To collision TTC. This research provides two ways of searching for the edge cases with Particle Swarm Optimization and Deep Q-learning. Notably, the paper focuses on a new and extensive method of finding edge cases. Particle Swarm Optimization is used to search over a larger amount of scenarios, whereas the Deep Q-Learning tunes each scenario to generate an edge case. This — to the best of our knowledge—has never been provided before in such a principled manner. When evaluating the results of the methods, both the algorithms outperform standard methods of grid search and randomized search by a factor of three and five, respectively.