The UTC Graduate School is pleased to announce that Viet Toan Tran will present Master’s research titled, Improving Traffic Management Efficiency through Reinforcement Learning-based Traffic Signal Control and Citywide Transit Simulation on 06/22/2023 at 09:30 in https://tennessee.zoom.us/j/95334444943 . Everyone is invited to attend.
Computer Science
Chair: Dr. Mina Sartipi
Co-Chair:
Abstract:
Traffic congestion presents a significant challenge, resulting in various negative outcomes such as decreased productivity, environmental harm, and adverse health effects due to stress. To alleviate this issue, two straightforward yet effective approaches involve enhancing traffic signal control and public transportation systems. However, prior research on this topic has limitations stemming from the absence of real-time reliable data. For instance, existing traffic signal controls often rely on inflexible fixed-phase plans or limited adaptability due to the constrained capacity of loop detectors. Similarly, studies related to public transit systems, such as designing, planning, and optimizing routes, require dependable citywide simulations that demand diverse and extensive data resources. Fortunately, the widespread adoption of new technologies such as cameras, Internet of Things (IoT) devices, and vehicular networks has significantly increased the collection speed and volume of traffic data, making real-time and reliable information more accessible. In this thesis, we explore how to leverage these data sources to enhance existing traffic signal controls (TSCs) and citywide public transportation simulations. For TSC, we introduce a comprehensive framework that facilitates the rapid prototyping of reinforcement learning (RL) methods, proposing a novel RL-based TSC employing world models. Additionally, we apply RL techniques to a digital twin and propose a new system that supports lifelong assessment. Regarding transit simulations, we develop a toolkit that calibrates citywide models using large-scale real-world observed data and propose a efficient way to address a simulation scenario involving changes solely in the transit system setting. Finally, we delve into the fundamental question of optimization for neural network training and propose a novel approach using neuroevolution, which surpasses the widely popular Gradient Descent method.