Deep spatio-temporal learning for dynamic urban shared mobility systems

Preview

PDF WRAP_Theses_Luo_2021.pdf - Submitted Version - Requires a PDF viewer. Download (21MB) | Preview

Request Changes to record.

Abstract

Shared mobility systems have been recently tested and piloted in many cities across the globe, with great potentials to be deeply woven into the fabric of future urban planning. Those systems, not only represent a more sustainable paradigm that can effectively cut unnecessary emissions, but could also bring significant societal benefits by offering a much more affordable on-demand mobility option to the general public. As a relatively new mobility trend, they expand at impressive speeds, pouring more fleet and infrastructure into the urban spaces than ever before. Coupled with the unpredictability of real-world environments, this brings numerous challenges in the deployment and operation of those systems, impacting their usability and practicality. In this thesis, we aim to better understand the structure, process and interaction of the shared mobility system with such dynamics, from a spatio-temporal learning perspective. Particularly, this thesis addresses the following question: How deep spatio-temporal learning can be tailored to improve the prediction, optimisation and actuation of shared mobility systems in the presence of real-world dynamics?

The key insight is that such dynamics, although complex, shouldn’t be treated independently in an isolated way, but rather be considered and embraced at full stack, from data modelling to inference and learning across the spatio-temporal domain. Specifically, we tackle this in a number of research threads. Firstly, we propose D3P, a novel data-driven prediction framework, which is able to directly model the spatio-temporal dynamics from the data with time-varying graphs, and uses bespoke dynamic Graph Convolutional Neural Networks (GCNs) to accurately forecast the future demand of the shared mobility systems in both short and long terms. We further proposeMANS, a new optimisation approach for mobility infrastructure deployment based on multi-agent neural search, which can effectively discover the optimal infrastructure deployment strategies for shared mobility systems across space and time, so that the services provided are ubiquitous to the users while sustainable in profitability. Finally, we propose ac-PPO, a novel user-incentive based fleet management approach, which uses a deep reinforcement learning paradigm to guide the rebalancing of fleet with cooperative users, and incorporates the dynamics by a new action cascading technique. All the proposed approaches have been comprehensively evaluated through large-scale experiments with extensive real-world datasets, and the results have shown superior performance compared with the state-of-the-art, demonstrating their potential impact on a broad spectrum of application scenarios.

Request changes or add full text files to a record

Repository staff actions (login required)

View Item