Repository logo
Log In(current)
  1. Home
  2. Colleges & Schools
  3. Graduate School
  4. Doctoral Dissertations
  5. Learning-Guided Design and Control for Urban Mobility
Details

Learning-Guided Design and Control for Urban Mobility

Date Issued
May 2026
Author(s)
Poudel, Bibek
The University of Tennessee
Advisor(s)
Weizi Li
Additional Advisor(s)
Sai Swaminathan
Fei Liu
Mariano Phielipp
Permanent URI
https://trace.tennessee.edu/handle/20.500.14382/57666
Abstract
Urban mobility is shaped by how transportation networks are designed and how traffic is controlled, and the interaction between these two problems governs what any system can achieve. Yet these problems are predominantly studied in isolation and under assumptions that abstract away variability in human behavior, non-stationary demand, and the interaction between infrastructure design and control. I study three problems that span this range: city-scale transit network design, mixed traffic control under realistic human driving, and joint optimization of crosswalk placement and signal timing on urban corridors. Across these problems, real-world demand and behavior data ground the learning problem, shaping both what the methods must handle and how they are evaluated. For city-scale transit network design, I introduce AlphaTransit, a framework that combines Monte Carlo Tree Search with a learned graph neural network to construct route networks with explicit lookahead. Transit route design is a combinatorial problem with a deceptive optimization landscape where locally appealing choices can steer the network toward globally poor configurations; AlphaTransit addresses this by scoring each route addition's long-term impact. On a realistic benchmark derived from Bloomington, Indiana, AlphaTransit achieves the highest service rate across two transit demand regimes. Under high transit demand, it achieves an 85.5 % service rate with under six-minute average wait time while using fewer buses than competing baselines. Under lower transit demand, it remains on the Pareto frontier of the service-rate/fleet-size trade-off. For mixed traffic control, I incorporate naturalistic driving data from the I-24 MOTION corridor and develop robot vehicle policies equipped with anticipatory congestion information. In the Ring environment, the safety-and-stability variant maintains time-to-collision above 4 s at all penetration rates, while the efficiency variant improves throughput by up to 54 % and fuel economy by up to 75 %. On urban corridors, I first demonstrate that adaptive signal control on the existing seven-crosswalk layout reduces pedestrian and vehicle wait times by up to 67 % and 52 %. I then introduce a co-optimization framework that jointly determines crosswalk placement and signal timing, reducing the layout from seven crosswalks to four while further cutting pedestrian and vehicle wait times by up to 79 % and 65 %. The co-optimized policy remains robust to post-training layout changes, achieving up to 97 % lower wait times than a sequentially trained controller. These results establish that design and control yield complementary gains that neither achieves in isolation. Taken together, this dissertation demonstrates that learning-guided optimization for urban mobility is most effective when matched to the intrinsic structure of each problem. Transit network design benefits from explicit lookahead over a combinatorial action space. Mixed traffic control benefits from naturalistic perturbations and anticipatory cues that reflect the variability of real human drivers. Urban corridors benefit from treating crosswalk placement and signal control as coupled decisions rather than separate stages. Across these problems, this perspective yields demand-responsive transit networks, proactive mixed-traffic control, and fewer, better-placed crosswalks that improve mobility for both pedestrians and vehicles.
Subjects

Computer Science

Reinforcement Learnin...

Transportation Engine...

Urban Mobility

Reinforcement Learnin...

Urban Mobility

Traffic Signal Contro...

Transit Network Desig...

Co-Optimization

Graph Neural Networks...

Degree
Doctor of Philosophy
Major
Computer Science
Embargo Date
May 1, 2027

University Libraries

1015 Volunteer Boulevard
Knoxville, TN 37996
865-974-4351

Map & Directions
Donate to the Libraries
  • About
  • John C. Hodges Society
  • Speaking Volumes magazine
  • Outreach
  • Directory
  • Employment
  • Policies
  • Library Intranet
University of Tennessee power T logo

The University of Tennessee, Knoxville
Knoxville, Tennessee 37996
865-974-1000

Events
A-Z
Apply
Privacy
Map
Directory
Give to UT
Accessibility

Built with DSpace-CRIS software - Extension maintained and optimized by 4Science