Abstract

Shared Mobility-on-Demand (MoD) has tremendously reshaped the transportation patterns in urban areas. The prosperity of Big Data and 5G network technology brings new challenges to shared MoD systems. Specifically, the major challenges in shared MoD systems under the data-driven environment include dynamic environment, stochasticity, large-scale optimization, and Big Data characteristics. In this research, we identify shared MoD management as a resource allocation problem in the two-sided market environment. Further, we adopt optimization under uncertainty modeling techniques to address the resource allocation issues for shared MoD systems. To improve the performance of the shared MoD systems under the data-driven environment considering uncertain parameters, we propose a generic learning-based data-driven optimization framework and apply it to three shared MoD optimization issues.

Specifically, we develop a generic learning-based data-driven optimization framework that integrates data processing, feature engineering, machine & deep learning approaches, reformulation, and decomposition algorithms to optimize the resource allocation problems in shared MoD systems. The research involves three synergistic shared MoD areas. (1) Kernel density enabled stochastic programming modeling for proactive vehicle allocation and reactive relocation in the car-sharing system. We design a data-driven optimization framework that seamlessly integrates kernel density estimation and a two-stage stochastic programming model to address the issue. (2) Mixture density networks enabled stochastic programming modeling for dynamic proactive idle vehicle guidance in the ride-hailing system. We design a novel deep neural network and integrate it with a one-stage stochastic programming model to guide the idle vehicle to alleviate the supply-demand gap in the ride-hailing region. (3) Dynamic data-driven robust optimization modeling for ride-sharing matching problems under travel time uncertainty. We develop a data-driven robust optimization framework that organically integrates gated recurrent networks, a one-stage robust optimization model, model reformulation, and a hybrid meta-heuristic algorithm to address the ride-sharing participants matching issue. Based on these research points, this Ph.D. thesis is a compilation of three published or submitted journal papers.

To sum up, this Ph.D. thesis presents a general data-driven optimization framework that organically integrates the artificial intelligence (AI) paradigm, including machine and deep learning approaches with optimization under uncertainty modeling techniques. Driven by AI techniques, the learning-based data-driven optimization framework can address the resource allocation problems in shared MoD systems by hedging against the uncertainty in a decent way. Further, the proposed approach is not limited to the shared MoD systems. Still, it can be applied to other relevant fields by replacing the learning and optimization components in the framework.