一句话：提出PWCC覆盖度，用运行剖面的概率质量加权t路场景特征组合，给出风险校准的测试充分性停止准则。
自动驾驶系统场景测试缺乏“何时停止”的判据：现有覆盖度只衡量结构多样性或组合广度，无法说明真实运行中还有多少预期条件未被测试。论文提出概率加权组合覆盖度（PWCC），从运行剖面数据学习预期运行条件的概率模型，为场景特征的t路组合分配概率质量，当未覆盖质量低于风险校准阈值时判定测试充分。在四个自动驾驶数据集上与七种基线对比，PWCC在DeepCollision上对故障检测预测最强（Spearman r_s=+0.981，比最佳组合基线高0.167且交互阶更低）；在DeepScenario上消除假充分，FSR=UFC=0，而CC平均UFC≥95.8%，并在其停止点暴露99.6%触发故障组合；得分对十倍蒙特卡洛样本量变化≤0.006且随套件规模单调增长。
Scenario-based testing of autonomous systems faces a fundamental unsolved problem: knowing when to stop. Existing coverage metrics measure either structural diversity or combinatorial breadth, but neither tells you how much of the system’s expected real-world operation remains untested. A test suite can score well on every existing metric while systematically missing the conditions the system will actually encounter. We introduce Probability-Weighted Combinatorial Coverage (PWCC), a metric that simultaneously measures coverage of the Op- erational Design Domain (ODD) and test adequacy. PWCC learns a probabilistic model of expected operating conditions from oper- ational profiling data, uses it to assign probability mass to 𝑡-way combinations of scenario features, and declares testing adequate when the uncovered mass falls below a risk-calibrated threshold. This gives practitioners something no existing metric provides: a principled, quantitative answer to “have we tested enough?” Evaluated on four autonomous driving datasets against seven baselines, PWCC demonstrates three properties. First, it is the strongest predictor of fault detection where the operational dis- tribution carries signal, reaching Spearman 𝑟𝑠 = +0.981 on Deep- Collision — 0.167 above the best combinatorial baseline at a lower interaction order. Second, it suppresses false sufficiency: on Deep- Scenario, PWCC achieves FSR = UFC = 0 across all splits versus CC’s mean UFC ≥ 95.8%, false sufficiency eliminated entirely, and, judged at its own stopping point, surfaces 99.6% of failure-triggering com- binations where combinatorial saturation stops after 6.4% of the campaign; on failure-rich datasets UFC drops by up to 4.9×. Third, it is stable: the score varies by at most 0.006 across a ten-fold increase in Monte Carlo sample size and grows monotonically with suite size. These results show that operational probability weighting is the decisive factor in avoiding dangerous false sufficiency verdicts; and that PWCC transforms coverage from a descriptive statistic into a risk-informed stopping decision.
@inproceedings{paperbot12442,
title = {PWCC: Probability-Weighted Combinatorial Coverage for Risk-Informed Scenario Testing of Autonomous Systems},
author = {Borhaneddine Hamadou and Chokri Mraidha and Nikolaos Matragkas},
booktitle = {ASE},
year = {2026},
doi = {10.1145/3832783.3837443}
}