McMaster University
Integrated Traffic and Infrastructure Management Using Deep Reinforcement Learning and Automated Vehicles
Abstract
dc:description.abstractTransportation systems are increasingly required to address multiple, and sometimes competing, management objectives, including congestion mitigation, traffic safety, emissions reduction, and infrastructure protection. These challenges are particularly important in emerging connected and automated vehicle environments, where new control opportunities are enabled by vehicle automation, vehicle-to-vehicle (V2V) communication, and vehicle-to-infrastructure (V2I) communication. However, conventional traffic management strategies are often limited in their ability to adapt to complex, nonlinear, and time-varying interactions between traffic operations and infrastructure constraints. For example, automated vehicle platooning can improve traffic flow efficiency and string stability, but closely spaced heavy-vehicle platoons may introduce bridge-loading concerns if infrastructure safety is not explicitly considered. Similarly, freeway bottlenecks and signalized intersections require adaptive control strategies that can respond to changing demand, vehicle composition, safety conditions, and environmental impacts. To address these gaps, this thesis develops and evaluates deep reinforcement learning (DRL)-based traffic control and automated vehicle management frameworks for integrated traffic and infrastructure management in connected and automated transportation environments. The thesis first develops a dynamic, bridge-safety-aware V2V truck platooning framework that reconciles freight efficiency with structural safety. The proposed framework adaptively modifies intra-platoon spacing as truck platoons approach bridges so that code-based bridge load constraints are satisfied while preserving free-flow throughput, platoon cohesion, and string-stable behavior. The thesis then investigates the application of DRL and automated vehicles (AVs) to freeway infrastructure management. A Rainbow-based DRL ramp metering controller is developed for a freeway corridor in which an upstream bridge speed restriction changes the traffic arriving at a downstream on-ramp merge. The controller uses a critical-speed reward that removes the need for a hand-selected weight between mainline conditions and ramp queues, and it is evaluated in microscopic traffic simulation across four bridge speed limits against a D3QN learner that shares the same reward, feedback-based PI-ALINEA metering, fixed-time metering, and no control. The results show that both reward-sharing learners hold mainline speeds well above fixed-time metering and no control, that the reward design rather than the choice of learner accounts for most of this gain, and that Rainbow learns faster and maintains shorter ramp queues than D3QN. A complementary analysis of unmetered traffic shows that increasing AV market penetration rates (MPRs) raise corridor speeds and progressively remove the differences between bridge speed limits. These freeway studies demonstrate how DRL and AVs can support infrastructure-aware traffic management, rather than focusing solely on mobility improvement. The thesis then extends this foundation by quantifying the joint impact of DRL-based traffic signal control and AVs on intersection efficiency, safety, and emissions across demand levels and AV MPRs. The findings demonstrate that DRL control significantly outperforms fixed-time and actuated baselines, and that its short-term benefits at low MPRs approach the long-term gains of conventional control at high MPRs. Building on this analysis, the thesis develops a multi-objective, platoon-aware deep reinforcement learning (MOPADRL) signal controller operating in a microscopic vehicle-to-everything (V2X) environment, in which V2V communication coordinates string-stable platoons and a conditional V2I platoon-request channel grants preemptive priority. MOPADRL jointly optimizes waiting time, CO2 emissions, and platoon cohesion, reducing queue length and waiting time by over 90\% relative to fixed-time control, lowering emissions by up to 47\% relative to human-driven baselines, and eliminating time-to-collision conflicts. The results further show that platoon protection and traffic efficiency function as cooperative, rather than antagonistic, objectives. Collectively, the four studies demonstrate that DRL and AVs can serve as complementary tools for next-generation transportation management. Overall, this thesis demonstrates that adaptive learning-based control can improve operational efficiency, safety, and emissions at intersections and freeway bottlenecks, while automated vehicle platooning can be managed in ways that protect both traffic flow performance and infrastructure safety. The findings provide practical insight into transitional deployment by showing how centralized control, decentralized AV behavior, and V2X-enabled platooning may interact as AV MPRs increase. From a broader perspective, this thesis contributes to the development of integrated traffic and infrastructure management systems that move beyond single-objective mobility optimization toward multi-objective, infrastructure-aware, and automation-ready control. Future research should extend the proposed frameworks to larger urban and freeway networks, incorporate mixed traffic with heterogeneous driver and vehicle automation behavior, consider uncertainty in communication reliability and sensing quality, and evaluate real-world implementation issues such as controller transferability, computational scalability, and coordination between traffic agencies and infrastructure owners.
Degree
thesis:*- Department dc:contributor.department
- Civil Engineering
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Karbasi, Amir Hossein
- Advisor dc:contributor.advisor
-
- Yang, Hao
Subjects
dc:subject × 5Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Identifier URI
- https://doi.org/10.71548/25627