Enhancing Underutilized Bus Routes with Advance Reservations and Semi-flexible Routing
Funded by: National Science Foundation (#2055347 )

We introduced a novel approach where bus riders provide advance notice of their stops, allowing buses to skip unrequested stops and take shortcuts. This method increases stop density, reduces walking distances to and from bus stops, and maintains operational efficiency. To design this system, we developed optimization models that maximize the number of stops while adhering to tour duration and arrival time constraints. A case study in Allegany County, Maryland, demonstrates significant enhancements for routes that were both underutilized and had layouts conducive to substantial shortcuts.

Paper:  https://doi.org/10.1287/trsc.2024.0561 

Empowering MetroAccess Service with Nested Decomposition and Service Type Integration
Funded by: National Science Foundation (#2055347 )

Through a university-industry collaboration, we have identified (1) modernizing the optimization engine in paratransit scheduling software suites, and (2) incorporating alternative service providers in paratransit service optimization as two critical steps in overcoming some challenges in paratransit practice. We developed a nested decomposition method in which a column generation–based solution approach is embedded in a temporal decomposition framework. Additionally, we integrated alternative service types, such as accessible taxis, in paratransit scheduling and designed a reoptimization procedure. The new optimization methods were implemented in a software suite of IT Curves and deployed in paratransit operations in the Washington, DC, metro area. It was found that the improved optimization engine, relative to the legacy, led to significant improvements in key operational metrics and yielded substantial operating cost reductions (approximately 15%). 

Paper:  https://doi.org/10.1287/inte.2024.0123 

Integrating dial-a-ride with transportation network companies for cost efficiency: A Maryland case study

Funded by: National Science Foundation (#2055347 )

In this paper, we explored outsourcing outlier trips to transportation network companies (TNCs) in Dial-a-ride (DAR) operations, using a two-stage solution framework based on mixed integer programming and multi-vehicle improvement heuristic, which achieved cost reductions of 7%-13% for a medium-sized DAR operator in Maryland. 

Paper:  https://doi.org/10.1016/j.tre.2023.103140   

Linear regression coupled Wasserstein generative adversarial network for direct demand modeling of ride-hailing trips in Chicago and Austin

While modern machine learning techniques enhance the predictive accuracy of zone-to-zone direct demand models, they often lack inference ability due to complex structures and are limited by small sample sizes of fixed origin-destination (OD) pairs. To address these issues, a linear regression coupled Wasserstein generative adversarial network (LR-WGAN) is proposed, combining linear explanatory components with non-linear patterns to improve prediction accuracy by generating additional OD data. This model has shown significant improvement in predicting OD ride-hailing demand in Chicago and Austin and allows for inference on the relationship between demand and predictor variables.  

Paper:  https://doi.org/10.1080/19427867.2024.2372944 

Gated ensemble of spatio-temporal mixture of experts for multi-task learning in ride-hailing system

Funded by: Miyan Research Institute, IUBAT

Designing spatio-temporal forecasting models separately in a task-wise and city-wise manner to forecast demand and supply-demand gap in a ride-hailing system poses a burden for the expanding transportation network companies. To address this issue, a deep multi-task learning architecture, GESME-Net, is proposed to simultaneously forecast demand and supply-demand gap across different tasks and cities. This framework employs a gated ensemble of spatio-temporal mixture of experts for convolutional recurrent neural network (CRNN), CNN, and recurrent neural network (RNN), as well as a task adaptation layer that reveals feature contributions while learning a joint representation for multi-task learning.

Paper: doi.org/10.1016/j.multra.2024.100166    

Using spatio-temporal deep learning for forecasting ride-hailing demand and supply-demand gap with anonymised spatial adjacency information

Funded by: Miyan Research Institute, IUBAT

Ride-hailing companies require accurate spatio-temporal forecasting of both demand and supply-demand gap to reduce passenger waiting times, minimize driver search friction, and ensure efficient service management. However, such forecasting is challenging due to strong spatio-temporal dependencies and, in many cases, the lack of explicit spatial adjacency information arising from unplanned cities and privacy issues. To address these issues, a spatio-temporal deep learning architecture integrating a feature importance layer, a one-dimensional convolutional neural network (CNN), and a zone-distributed independently recurrent neural network (IndRNN) is introduced to handle anonymized spatial data. 

Paper: doi.org/10.1049/itr2.12073