Date of Award
Doctor of Philosophy in Transportation - (Ph.D.)
Civil and Environmental Engineering
I-Jy Steven Chien
This study proposes a Decentralized Sparse Coordination Learning System (DSCLS) based on Deep Reinforcement Learning (DRL) to control intersections under the Connected and Automated Vehicles (CAVs) environment. In this approach, roadway sections are divided into small areas; vehicles try to reserve their desired area ahead of time, based on having a common desired area with other CAVs; the vehicles would be in an independent or coordinated state. Individual CAVs are set accountable for decision-making at each step in both coordinated and independent states. In the training process, CAVs learn to minimize the overall delay at the intersection. Due to the chain impact of taking random actions in the training course, the trained model can deal with unprecedented volume circumstances, the main challenge in intersection management. Application of the model to a single-lane intersection with no turning movement as a proof-of-concept test reveals noticeable improvements in traffic measures compared to three other intersection control systems.
A Spring Mass Damper (SMD) model is developed to control platooning behavior of CAVs. In the SMD model, each vehicle is assumed as a mass, coupled with its preceding vehicle with a spring and a damper. The spring constant and damper coefficient control the interaction between vehicles. Limitations on communication range and the number of vehicles in each platoon are applied in this model, and the SMD model controls intra-platoon and inter-platoon interactions. The simulation result for a regular highway section reveals that the proposed platooning algorithm increases the maximum throughput by 29% and 63% under 50% and 100% market penetration rate of CAVs. A merging section with different volume combinations on the main section and merging section and different market penetration rates of CAVs is also modeled to test inter-platoon spacing performance in accommodating merging vehicles. Noticeable travel time reduction is observed in both mainline and merging lanes and under all volume combinations in 80% and higher MPR of CAVs.
For a more reliable assessment of the DSCLS, the model is applied to a more realistic intersection, including three approaching lanes in each direction and turning movements. The proposed algorithm decreases delay by 58%, 19%, and 13% in moderate, high, and extreme volume regimes, improving travel time accordingly. Comparison of safety measures reveals 28% improvement in Post Encroachment Time (PET) in the extreme volume regime and minor improvements in high and moderate volume regimes. Due to the limited acceleration and deceleration rates, the proposed model does not show a better performance in environmental measures, including fuel consumption and CO2 emission, compared to the conventional control systems. However, the DSCLS noticeably outperforms the other pixel-reservation counterpart control system, with limited acceleration and deceleration rates. The application of the model to a corridor of four interactions shows the same trends in traffic, safety, and environmental measures as the single intersection experiment.
An automated intersection control system for platooning CAVs is developed by combining the two proposed models, which remarkably improves traffic and safety measures, specifically in extreme volume regimes compared to the regular DSCLS model.
Mirbakhsh, Ardeshir, "A self-learning intersection control system for connected and automated vehicles" (2022). Dissertations. 1605.