An accelerated black-box sample paths feasibility probability estimation method for control policies of autonomous vehicles

With the continuous development of autonomous vehicles technology, the feasibility probability estimation of the sample paths has become a key requirement in the performance evaluation of its control policies and the policy optimization with chance constraints. Aiming at the defect that current autonomous driving testing methods generally rely on human prior knowledge for sampling allocation, this paper proposes a method that can allocate the number of samples according to the feasibility probability and state occurrence probability, and proves its optimality. In this paper, we first propose an optimal sampling times allocation method to minimize probabilistic estimation variance, which can obtain an acceleration effect that is reciprocal to the probability of occurrence of the most critical state. For the actual task requirement, we also propose algorithms with iterative estimation and low-fidelity models. The results from numerical experiments with two initial states and intelligent vehicle cornering cruise experiments under ten initial states demonstrate that our method can achieve the same prediction estimation error with fewer samples.