Using Machine Learning to Improve Mobility-on-Demand Systems Integrated into Public Transit

• Grant number: 23/17501-4
• Support Opportunities: Scholarships abroad - Research
• Start date: March 30, 2025
• End date: March 29, 2026
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Raphael Yokoingawa de Camargo
• Grantee: Raphael Yokoingawa de Camargo
• Host Investigator: Amer Shalaby
• Host Institution: Centro de Matemática, Computação e Cognição (CMCC). Universidade Federal do ABC (UFABC). Ministério da Educação (Brasil). Santo André , SP, Brazil
• Institution abroad: University of Toronto (U of T), Canada
• Associated research grant: 21/11959-3 - CITIES: Center for Innovation in Urban Public Policies, AP.CCD

Abstract

The evolving landscape of urban mobility has witnessed the advent of Mobility-on-Demand (MoD) systems, including ride-hailing services. There is increasing interest in applying this concept to public transportation, for instance, for first- and last-mile connectivity to public transportation hubs through microtransit solutions like on-demand vans. However, dynamically managing these systems poses significant challenges due to varying demand patterns, uncertain traffic conditions, and the need for effective vehicle assignment, routing, and fleet rebalancing.In this project, we will propose and evaluate Deep RL algorithms for selecting target routes for microtransit in the first- and last-mile scenarios with connections to public transit hubs. The selected target routes will implicitly perform the vehicle assignment and rebalancing steps. There are several studies on the use of Deep RL for ride-hailing, but they focus mostly on the repositioning problem only and focus on the scenario of door-to-door service. Also, the algorithms are trained on a single region, and to work on different regions, they need retraining. This project will generate three contributions to the existing literature: 1) a model of Deep RL for the context of first- and last-mile transportation to and from transportation hubs using microtransit; 2) a transferable state representation for Deep RL, facilitating model training using data from multiple regions; and 3) a hierarchical state representation Deep RL models for larger regions containing multiple transportation hubs. We will evaluate the algorithms using passenger-centric and operational metrics, comparing them to existing vehicle assignment, routing, and fleet rebalancing algorithms.