Core Concepts
The author proposes a method using dense reinforcement learning to optimize combination coefficients for adaptive testing of connected and automated vehicles, enhancing evaluation efficiency.
Abstract
The content discusses the challenges in evaluating safety performance for connected and automated vehicles (CAVs) and introduces an adaptive testing environment generation method. The approach involves dense reinforcement learning to optimize combination coefficients, improving evaluation robustness. The paper presents theoretical analysis, results from high-dimensional overtaking scenarios, and comparisons between different methods like NDE, NADE, and AdaTE.
The assessment of safety performance is crucial for CAVs' development and deployment. Existing studies focus on adaptive testing scenarios due to differences between CAVs and prior knowledge. Adaptive testing methods aim to generate scenarios dynamically during evaluation processes but face challenges with high-dimensional scenarios due to rarity and dimensionality issues. The proposed AdaTE method enhances evaluation robustness by optimizing combination coefficients using dense reinforcement learning.
To address the limitations of existing methods, the authors introduce an adaptive policy based on surrogate-to-real gaps for critical state-action pairs. The DenseRL method efficiently learns regression targets by focusing on critical scenes showing significant surrogate-to-real gaps. Results demonstrate that AdaTE achieves higher evaluation efficiency compared to NDE and NADE while ensuring evaluation robustness.
Stats
To increase evalu-ation efficiency, recent years have seen rapid advancements in the field of testing scenario library generation.
Overtaking scenarios will exceed 1400 dimensions.
For AV-I, AV-II, and AV-III, the required number of tests to reach the ASD threshold are 3.8 × 104, 4.9 × 104, and 3.5 × 104 respectively.
The optimized combination coefficients for AV-I are [0.95, 0.03, 0.02], for AV-II are [0.82, 0.16, 0.02], and for AV-III are [0.65, 0.02, 0.33].
Quotes
"The assessment of safety performance plays a pivotal role in the development and deployment of connected and automated vehicles (CAVs)."
"To improve evaluation efficiency, we optimize the combination coefficients of SMs through adaptive testing."
"Our approach focuses on learning the values of critical state-action pairs exhibiting significant surrogate-to-real gaps."