Multitask Learning and Prediction of Baseline Driving Performance Measures

Driving performance measures (DPMs) are important indices for driving and personal safety in vehicle operation. The DPMs are collected under various controlled driving conditions to demonstrate different driving behaviors so that mitigating technology interventions can be studied and designed. However, significant costs are involved in the DPM acquisition, and there are a very limited number of controlled driving condition data. Thus, the modeling and prediction of the DPMs under unobserved driving conditions are critical, and many methods have been developed. However, existing literature in this area suffer a common limitation: The interactions among different DPMs are not fully considered (each DPM is modeled individually), although the existence of such interactions is widely reported. This paper proposes a novel DPM modeling and prediction method, i.e., multi-output convolutional Gaussian process (MCGP), that incorporates the interactions among different DPMs. The method features the modeling flexibility for different DPMs and the interpretable modeling structure for integrating the DPM interactions. The method is compared with three benchmark methods on the DPM data set under four different settings, and the results demonstrate the superiorities of the method. Discussions and interpretations of the results are also provided.驾驶性能指标（Driving Performance Measures, DPMs）是车辆运行中关乎驾驶与人身安全的重要指标。该类指标会在多种受控驾驶工况下采集，以表征不同的驾驶行为，进而为缓解类技术干预手段的研究与设计提供支撑。然而，DPM的采集需耗费高昂成本，且现有受控驾驶工况数据的体量极为有限。因此，针对未观测驾驶工况下的DPM建模与预测研究至关重要，目前已涌现出诸多相关方法。但该领域现有研究普遍存在一项共性局限：尽管已有大量研究证实不同DPM间存在交互效应，但现有方法并未充分考量这类交互，而是仅对单个DPM进行独立建模。为此，本文提出一种新型DPM建模与预测方法——多输出卷积高斯过程（Multi-output Convolutional Gaussian Process, MCGP），该方法能够整合不同DPM间的交互效应。该方法兼具针对不同DPM的建模灵活性，以及用于整合DPM交互效应的可解释建模结构。本文在DPM数据集上，针对四种不同实验设置，将所提方法与三种基准方法进行对比，实验结果证实了所提方法的性能优势。此外，本文还对实验结果展开了讨论与解读。