By Seyedehsima Madani
Hybrid transit systems offer a promising solution for improving urban accessibility and public transportation efficiency by combining the reliability of fixed-route transit (FRT) with the flexibility of demand-responsive transit (DRT). However, these services are often planned independently, overlooking the strong interdependencies between network design and on-demand operations. This dissertation addresses this challenge by developing integrated optimization and machine learning approaches to support the efficient and coordinated design of hybrid public transportation systems.
The first part of the thesis presents a systematic literature review of 83 studies on the integration of fixed-route and demand-responsive transit systems. It introduces a unified classification framework based on the degree of integration, planning level, and optimization perspective. The review identifies important research gaps, particularly the limited attention given to the joint strategic design of fully integrated hybrid transit systems.
Building on these findings, the second part introduces a bi-level optimization framework that jointly considers fixed-route network design and demand-responsive transit operations. The framework combines Particle Swarm Optimization (PSO) for designing fixed-route networks, including lines, frequencies, and timetables, with Large Neighborhood Search (LNS) for optimizing demand-responsive vehicle routing and scheduling. Computational experiments demonstrate that joint planning can reduce total fleet requirements by approximately 11% compared to conventional sequential approaches, while maintaining comparable service coverage and passenger travel times.
The framework also examines how different passenger preferences, service coverage requirements, and spatial demand distributions influence optimal transit network configurations. The findings demonstrate that integrated planning enables more efficient allocation of resources between fixed-route and demand-responsive services, allowing transit systems to adapt to diverse mobility needs and operating conditions.
Finally, the dissertation introduces a machine learning–enhanced optimization approach to address the computational complexity of integrated transit design. Different surrogate models, including traditional machine learning methods and graph neural networks, are evaluated to approximate computationally expensive operational decisions. XGBoost demonstrates the strongest predictive performance, enabling substantial reductions in computational time. A hybrid strategy combining surrogate predictions with periodic explicit optimization evaluations further improves reliability, achieving a 4–6 times reduction in runtime while maintaining high solution quality.
Together, these contributions demonstrate the potential of combining mathematical optimization and machine learning to develop more efficient and scalable planning approaches for hybrid transit systems. By bridging strategic network design and operational planning within a unified decision-support framework, this dissertation advances both methodological research and practical transportation planning. The findings provide valuable insights for transport authorities, policymakers, and researchers seeking to develop more accessible, flexible, and sustainable public transportation systems.