Accelerated Testing and Evaluation for Black-Box Autonomous Driving Systems via Adaptive Markov Chain Monte Carlo

Black-box decision-making algorithms based on deep learning and reinforcement learning have demonstrated significant advancements in autonomous vehicles (AVs), but still encounter challenges due to unpredictable failures in long-tail scenarios. The efficient evaluation of the safety of these algorithms is essential for the further development of AVs. This paper introduces a novel testing and evaluation method that leverages subset simulation (SuS) with adaptive Markov chain Monte Carlo (ApMCMC). The method is designed to facilitate the occurrence of rare events and expedite the probability assessment process, thereby enabling rapid evaluation of the failure probability of deep reinforcement learning (DRL)–based end-to-end decision-making algorithms. The car-following scenario is used as a case study to demonstrate the efficacy of our proposed method and four state-of-the-art DRL-based algorithms are trained as evaluation objects. Experimental results indicate that our method maintains evaluation errors within 5%. Compared with the naive Monte Carlo method and other existing evaluation techniques, our approach significantly enhances evaluation accuracy, reduces variance, and improves confidence levels, while accelerating evaluation efficiency by 5-200 times. Additionally, our method can easily handle end-to-end black-box decision-making systems and exhibits the adaptability to be seamlessly extended and applied to other scenarios.