Predictive Modeling of Autonomous Vehicle Behavior With Imbalanced & Mixed Data

Autonomous vehicles are equipped with proprietary Artificial Intelligence (AI) technology, which is confidential and inaccessible. To ensure its operational integrity predective models are used. These models integrate numerous data sources and machine learning techniques to enhance the vehicle’s ability to navigate, recognize objects, and avoid obstacles safely and efficiently. Before being marketed in Europe, vehicles undergo a series of tests conducted in diffrentes scenarios. When the decision made by vehicle corresponds to a safe situation, the test result is considered as OK. Otherwise, it is KO. Hence, this can be seen as a classification problem and can be modeled by implementing a classifiers trained on observed data. However, it has been observed that datasets contain imbalanced data, with a higher number of OK tests compared to KO trials. To implement the classifier effectively, this imbalance must be addressed. Various solutions from the literature, including resampling techniques like SMOTE and its variants adapted for mixed data, can help. This study aims to apply these methods to a real dataset derived from vehicle tests incorporating AI. The SMOTE-ENC variant has demonstrated its effectiveness, particularly in classifying the positive KO class, correctly identifying 5 out of 6 tests. In contrast, the non-rebalanced approach misclassifies 5 out of 6 tests.