Online Personalized Preference Learning Method Based on In-Formative Query for Lane Centering Control Trajectory

• Published in: Sensors (Basel, Switzerland) Vol. 23; no. 11; p. 5246
• Main Authors: Ran, Wei, Chen, Hui, Xia, Taokai, Nishimura, Yosuke, Guo, Chaopeng, Yin, Youyu
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI AG 31.05.2023; MDPI
• Subjects: Autonomous vehicles; Bayesian approach; Behavior; Customization; Driverless cars; Efficiency; Expected utility; LCC trajectory; Machine learning; Methods; online learning; preference learning; Preferences; Queries; Robots; User feedback; utility theory; United States > US; Bayesian approach; LCC trajectory; online learning; preference learning; utility theory
• DOI: 10.3390/s23115246
• ISSN: 1424-8220

The personalization of autonomous vehicles or advanced driver assistance systems has been a widely researched topic, with many proposals aiming to achieve human-like or driver-imitating methods. However, these approaches rely on an implicit assumption that all drivers prefer the vehicle to drive like themselves, which may not hold true for all drivers. To address this issue, this study proposes an online personalized preference learning method (OPPLM) that utilizes a pairwise comparison group preference query and the Bayesian approach. The proposed OPPLM adopts a two-layer hierarchical structure model based on utility theory to represent driver preferences on the trajectory. To improve the accuracy of learning, the uncertainty of driver query answers is modeled. In addition, informative query and greedy query selection methods are used to improve learning speed. To determine when the driver’s preferred trajectory has been found, a convergence criterion is proposed. To evaluate the effectiveness of the OPPLM, a user study is conducted to learn the driver’s preferred trajectory in the curve of the lane centering control (LCC) system. The results show that the OPPLM can converge quickly, requiring only about 11 queries on average. Moreover, it accurately learned the driver’s favorite trajectory, and the estimated utility of the driver preference model is highly consistent with the subject evaluation score.