Chaeeun Han |

SAGE: Similarity-Based Cleaning of Poisoned Training Data from Verified Examples

Sun, 01 Nov 2026 00:00:00 +0000

An End-to-End Solution for Public Transit Stationing and Dispatch Problem

Sat, 01 Nov 2025 00:00:00 +0000

Fixed-Line Transit 2.0: Real-Time Optimization of High-Frequency Transit Service

Tue, 01 Jul 2025 00:00:00 +0000

This project will develop and deploy a real-time, AI-driven decision support system that makes high-frequency bus service more reliable by anticipating and proactively mitigating delays, overcrowding, and service disruptions. Frequent, high-ridership bus routes are inherently unstable: when a bus falls behind, it encounters more waiting passengers and falls even further behind, until buses bunch together and leave long gaps in service, causing longer and less predictable waits, overcrowding, and missed transfers. The project focuses on WeGo Public Transit’s Route 55 Murfreesboro in Nashville, Tennessee, WeGo’s busiest route with nearly 5,000 weekday boardings, but also its least reliable one: more than 1 in 4 buses are either bunched or gapped, adding up to over 105 hours of excess waiting time for passengers every weekday. The project is led by WeGo in partnership with the Nashville Department of Transportation and Multimodal Infrastructure, the Pennsylvania State University, Vanderbilt University, and technology companies.

At the heart of the project is an Integrated Decision Support System that fuses heterogeneous real-time data, such as bus locations updated every two seconds, passenger counts, traffic, and weather. Computer vision will count passengers at bus stops and on buses, and machine learning models will forecast bus arrival times, headways, overcrowding, and service disruptions. Building on these forecasts, the system will warn dispatchers of emerging problems and use AI-driven optimization to recommend strategies that prevent them, such as holding or short-turning buses, skipping stops, or inserting standby buses. Adaptive traffic signal control with headway-conditional transit signal priority, continuously optimized using machine learning, will further reduce travel times and their variability, and riders will receive more accurate arrival predictions in their smartphone apps and on digital displays at bus stops.

Overview of the Fixed-Line Transit 2.0 system

Together with Vanderbilt University, our team at Penn State is responsible for the AI at the core of this system: prediction algorithms for bus arrivals and headways, and a real-time decision support system for dispatchers and drivers. We will create a decision-focused learning framework, in which prediction models trained on spatiotemporal data, such as weather, traffic, and ridership, are incorporated into non-myopic online optimization algorithms for proactive headway management and vehicle stationing. These algorithms will use the learned models to simulate uncertain future trajectories of the transit system and reason through them to improve service reliability. Our approach will combine these strengths of AI with the strengths of human supervisors, whose tacit knowledge and expertise are invaluable for understanding the dynamic challenges of fixed-line transit, while giving supervisors, drivers, and passengers real-time situational awareness of the whole system.

By making frequent bus service faster and more predictable, the project will improve access to jobs, education, and essential services: 65% of Route 55 riders do not have access to a vehicle, and since the route has more transfers to and from other routes than any other WeGo route, the benefits will extend to residents in many other parts of Nashville. More reliable transit may also draw people away from driving, and adaptive signals with transit signal priority will reduce delays at intersections, together reducing congestion and improving energy efficiency. The decision support system will also improve the productivity and job quality of dispatchers and drivers. Unlike dedicated transit lanes, whose construction is a multi-year, resource-intensive undertaking, the system builds on existing infrastructure, such as municipal fiber, onboard sensors, and real-time transit data feeds, greatly reducing the cost, duration, and risk of deployment. Extending it to six additional high-frequency WeGo routes would bring more than 2 in 3 WeGo bus boardings, nearly 20,000 rides per weekday, onto “Transit 2.0” routes, and the approach is relevant to other transit agencies facing similar challenges.

An Online Approach to Solving Public Transit Stationing and Dispatch Problem

Wed, 01 May 2024 00:00:00 +0000

Acceptance Rate: 23%

Forecasting and Mitigating Disruptions in Public Bus Transit Services

Wed, 01 May 2024 00:00:00 +0000

Acceptance Rate: 25%

Microtransit Solutions for Underserved Communities

Sat, 01 Aug 2020 00:00:00 +0000

Public transportation infrastructure is an essential component in cultivating equitable communities. However, public transit agencies have historically struggled to achieve this since they are often severely stressed in terms of resources as they have to make the trade-off between concentrating service into routes that serve large numbers of people and spreading service out to ensure that people everywhere have access to at least some service. A solution that holds great promise for improving public transit systems is the integration of fixed-route services with microtransit systems: multi-passenger transportation services that serve passengers using dynamically generated routes and may expect passengers to make their way to and from common pick-up or drop-off points. However, most microtransit systems have failed in the past due to the lack of community engagement, inability to handle the uncertainty of operations when integrating the fixed transit, and inability to handle the system-level optimization challenges.

This project takes a socio-relational approach to community engagement in collaboration with the Chattanooga Area Regional Transportation Authority (CARTA), designs a community-centric micro-transit service that augments fixed-line public transit networks (improving transit accessibility), and demonstrate its effectiveness in the representative city of Chattanooga. The outcome of the project will be a deployment-ready software system that can be used by an agency to design and operate a micro-transit service effectively. The algorithmic toolchain will be complemented by mechanisms to optimally select the parameters and sustainably manage the data required by the algorithms. In addition, this project will provide a set of exemplar case studies and a validated social methodology to engage the community and learn their requirements, which will be fed into the algorithms. This will potentially impact a wide range of cities in the U.S. that do not have well-developed transit systems as the project will not only provide a reusable operations system but also demonstrate how integrated socio-technical research and strong community engagement can provide a pattern for sustainability and expansion.

The intellectual merit of this project lies in the novel community engagement approach and combined operations research and data-driven, learning-based integrated system optimization. Towards this goal, the project will make four key contributions.

First, the project will develop a novel targeted outreach approach that uses the relational networks of social capital (e.g., outreach to community centers, congregations and faith communities, schools, and similar organizational structures) and builds a categorical demand model to design an innovative micro-transit system. The project’s hypothesis is that the behavioral impact on public-transit ridership with the proposed method will be significantly higher than with an approach that focuses only on the economic or time-saving benefits of the improved transit system.
Second, the project will introduce a sustainable and resilient data-integration platform that dynamically adjusts the location of the sensor data used to affect the design parameters and assess performance of the transit system. This is crucial because cloud computing is still very expensive for community partners, especially for real-time high-velocity and high-volume data analysis. Further, this data store provides us an opportunity to take a privacy-by-design approach for the datasets collected during the project. Specifically, the project will develop novel integrated anonymization mixers for multi-modal datasets (e.g., mixing information of different modalities, such as location traces and transactions, together in a spatiotemporal-transactional mixer) that achieve a given level of privacy (quantified using the notation of differential privacy) while maximizing the accuracy of transit queries, relying on not just the privacy-properties of the data but also on the needs of the queries.
Third, the project will develop uncertainty-aware fleet management and dispatch algorithms that incorporate demand aggregation and environmental uncertainty caused by congestion, incidents, and their impact on the system (both for the users and fleet operations).
Fourth, the project will leverage recent advances in active learning for non-stationary environments with contextual side information to design algorithms that will aid in the exploration and optimal selection of hyperparameters for microtransit algorithms.