Feature Reduction for Scenario-Based Safety Assessment of Automated Vehicles

Abstract

Automated Vehicles (AVs) have the potential to significantly reduce road crashes. To enable widespread adoption of AVs, rigorous safety assessment is essential. Valuable insights for AV safety validation can be gained by simulating scenarios across a broad range of feature variations. A common challenge in simulating these scenarios is known as the curse of dimensionality, where increasing the number of scenario features requires a near-infinite number of simulations to cover all variations. This issue of complexity presents a need for reducing scenario features. Most related work focuses on identifying important scenario features, while few evaluate how reducing these features impacts AV failure estimation. This study aims to address this gap by evaluating a wide range of feature reduction methods and assessing their effect on AV failure estimation. Previous research generated datasets for three distinct scenario categories by performing simulations using driver reference models on real-world data. The machine learning classifiers XGBoost and random forest were applied to this data for predicting AV failures, from which XGBoost achieved best performance. Ten dimensionality reduction techniques, including both feature selection and transformation approaches, were employed to reduce the scenario feature set. The optimal reduced feature set was selected based on Area Under the Precision Recall Curve (AUPRC) classification performance. To assess the impact on AV failure estimation, results were compared against those obtained with the full set of features. The findings indicate that reliable AV failure estimates can be maintained, and even marginally improved, after reducing the data to a small fraction of the original feature set.