Koopman-Based Data-Driven Techniques for Adaptive Cruise Control System Identification

Accurately identifying the intrinsic model of Adaptive Cruise Control has the potential to enhance the prediction of automated car-following behavior, helping vehicles' decision-making and contributing to safer and more efficient traffic flows. Moreover, white box models offer an analytical base for evaluating the impact of automated driving functions on macroscopic traffic dynamics, consequently aiding the management of the whole intelligent transportation system. Many existing system identification techniques have been applied to automated vehicles. However, most of these studies focus on identifying parameters for models of a fixed prototype. Their reliance on accurate estimation of state time derivatives prevents their real applications, challenged by low sampling rates, noisy measurements, and limited observation periods. In contrast, the Koopman operator learning framework presents a promising improvement that can identify the nonlinear evolutionary properties of continuous-time systems. In this study, we apply Koopman-based methods to data driven Adaptive Cruise Control model identification. Additionally, as the challenge remains in establishing a practical relationship between identification accuracy and sampling rate, we numerically compared the performance of three Koopman-based learning frameworks, finite-difference, Koopman-logarithm, and a newly devised resolvent-type method, with that of a commonly used offline simulation-based batch optimization approach. We introduce a novel modification to the resolvent-type method, and the experimental results demonstrate its state of the art performance, particularly in identifying the potential existence of parametric noise at lower sampling rates.