Reinforcement Learning for Preventing Bus Bunching in UVA Transit: A Proposal for Real-Time Holding Control and Dispatch Decision Support/A Meta-Study of Task-Level Utility in Differentially Private Census Data Legitimacy, Governance, and the U.S. Census Bureau’s Deployment of Differential Privacy 6 views

Modern infrastructure depends on technical systems that must make decisions under uncertainty while remaining useful to the people who rely on them. In my capstone project, "Reinforcement Learning for Preventing Bus Bunching in UVA Transit”, I examine reinforcement learning applications in supporting real-time dispatch decisions in the University of Virginia (UVA) transit system. Bus bunching creates longer waits, uneven spacing, and unpredictable arrivals for students, staff, and visitors; dispatchers must respond to changing traffic and passenger demand in real time. In my STS research paper, "Legitimacy, Governance, and The U.S. Census Bureau’s Deployment of Differential Privacy, I examine how the Census Bureau’s 2020 deployment of the differential privacy mechanism in the Centennial Census shaped the legitimacy and practical utility of census data across stakeholder communities. Census statistics guide representation, funding, public health analysis, and planning so changes to how these statistics are protected and released carry both technical and institutional consequences. Both of these projects are connected by their focus on how computational systems should be evaluated through the tasks they support. In transit control and census data protection, the fundamental question is how a technical system can modulate uncertainty without weakening the ability of affected communities to act on its outputs. The expected conclusion of the capstone project is that reinforcement learning can provide an implementable decision support framework for improving transit regularity in small or medium scale transit systems. The project does not require full automation of dispatching decisions. It proposes a model that can produce consistent, statistically-backed recommendations that leaves operational authority with human dispatchers. In simulation, the expected result is that buses remain more evenly spaced across a route. This improvement should be most evident when passenger loads, travel times, or weather conditions disrupt typical operations. The system would reduce patterns where one bus becomes overcrowded and delayed while the following bus runs under lighter demand. More generally, the project argues that reinforcement learning is valuable when operational decisions have delayed consequences and competing objectives. A holding decision may improve spacing at one point while a downstream delay is generated later; the system must learn policies that incorporate long term effects instead of local conditions alone. The paper concludes that legitimacy challenges primarily emerged from how privacy and utility tradeoffs were structured and communicated under Bureau centered governance. It recommends a plan for the 2030 deployment that would require explicit stakeholder engagement, more transparency, independent review structures that tie technical performance with accountability.

BS (Bachelor of Science)

School of Engineering and Applied Science Bachelor of Science in Computer Science Technical Advisor: Rosanne Vrugtman STS Advisor: Pedro Francisco Technical Team Members: David Antrobius

All rights reserved by the author (no additional license for public reuse)

2026-05-09