| dc.description.abstract | This dissertation focuses on computational approaches for the design and implementation of real-time transportation systems. We focus on demand-responsive transportation, which can be thought of as any transit solution that adapts to changes in demand, the environment and resources in real-time. The design of demand-responsive transit requires solving complex combinatorial optimization problems where decisions are made over large geographies with computationally intractable state-action spaces. Exact solutions are infeasible in this setting. We address this problem through data-driven methods which utilize vast sums of data generated from the connected devices now ubiquitous in urban areas.
We propose SmartTransit-AI: A computational framework for demand-responsive transportation. Our framework models the problem as a real-time resource allocation problem that can be decomposed into four concrete sub-problems - 1) Planning, 2) Prediction, 3) Deployment, and 4) Software Design. We make contributions towards three critical areas. First, we propose a non-myopic, adaptive planning algorithm for ride-pooling. Second, we use multi-task and transfer learning to overcome data quality issues when designing energy prediction models for fixed-line services. Third, we resolve challenges related to anomaly detection and application design for edge-network deployments in the transportation domain.
The results of this project have been codified into a cloud application for managing
and deploying demand-responsive transportation operations. The application framework targets transit agencies and allows them to manage their fleets, schedules and bookings through a collection of web and mobile interfaces. The software exposes the state-of-the-art optimization algorithms and tools proposed in this work with the goal of easing technology transfer between research and practice. Lastly, we detail our experiences deploying the framework with our partner transit agency in a real-world setting. | |
