Extreme Scenario Generation Based on Adversarial Attack

Abstract

The transportation field requires a large number of simulation scenarios for testing. At present, there is relatively little research on the generation of extreme scenarios. In this paper, we give the definition of extreme scenarios, which are prone to problems, and divide them into two categories: the extreme scenarios based on primitive value and the extreme scenarios based on primitive coupling. This paper focuses on the second which considers the coupling effect of different primitives in the scenarios, using the methods of adversarial attack: FGSM, FGSM-target, BIM, ILCM, PGD and strategically-timed attack. Using vehicle agent for test, the first five methods prove the feasibility and effectiveness of extreme scenario generation, and the sixth method simplifies the generation process.