Trajectory Based Traffic Analysis and Control Utilizing Connected Autonomous Vehicles

Recent scholars have developed a number of stochastic car-following models that have successfully captured driver behavior uncertainties and reproduced stochastic traffic oscillation propagation. While elegant frequency domain analytical methods are available for stability analysis of classic deterministic linear car-following models, there lacks an analytical method for quantifying the stability performance of their peer stochastic models and theoretically proving oscillation features observed in the real world. To fill this methodological gap, this study proposes a novel analytical method that measures traffic oscillation magnitudes and reveals oscillation characteristics of stochastic linear car-following models. We investigate a general class of stochastic linear car-following models that contain a linear car-following model and a stochastic noise term. Based on frequency domain analysis tools (e.g., Z-transform) and stochastic process theories, we propose analytical formulations for quantifying the expected speed variances of a stream of vehicles following one another according to one such stochastic car-following model, where the lead vehicle is subject to certain random perturbations. Our analysis on the homogeneous case (where all vehicles are identical) reveals two significant phenomena consistent with recent observations of traffic oscillation growth patterns from field experimental data: A linear stochastic car-following model with common parameter settings yields (i) concave growth of the speed oscillation magnitudes and (ii) reduction of oscillation frequency as oscillation propagates upstream. Numerical studies verify the universal soundness of the proposed analytical approach for both homogeneous and heterogeneous traffic scenarios, and both asymptotically stable and unstable underlying systems, as well as draw insights into traffic oscillation properties of a number of commonly used car-following models. Overall, the proposed method, as a stochastic peer, complements the traditional frequency-domain analysis method for deterministic car-following models, and can be potentially used to investigate stability responses and mitigate traffic oscillation for various car-following behaviors with stochastic components. Emerging connected and autonomous vehicle (CAV) technologies enable the accurate implementation of vehicle trajectory optimization algorithms. An adaptive trajectory controller is demanded to overcome the heterogeneity of CAV technologies. We develop a dynamic data-driven control architecture to optimally control CAV trajectories. For the sake of short term planning in transportation, we currently focus on mixed traffic scenarios and individual CAV control. The controller utilizes real-time spatio-temporal information to provide better solutions compared to classical controllers which only capture features of specific data with fixed parameters or suffer from model errors of mathematical formulations. A reinforcement learning method is applied and a triple-thread structure is proposed to take advantage of dynamic learning and guarantee driving safety at the same time.The present study then proceeds with an illustration with empirical data collected in field tests. The proposed controller is shown to outperform human drivers and Adaptive Cruise Control methods. Impact of the size of spatial information and temporal information, as well as changes of reward functions on the controller learning speed and performance are discussed