Data-driven management of urban mobility: spatiotemporal prediction and learning-based adaptive control
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
The increasing density of urban environments presents significant challenges to transportation systems, demanding more intelligent and adaptive management strategies. While data-driven methods offer a promising paradigm, existing artificial intelligence (AI) models often rely on overly simplistic assumptions, limiting their effectiveness in complex, real-world scenarios. For instance, many models fail to capture the intricate spatiotemporal dependencies in the data or oversimplify scenarios as fully observable. Moreover, some incomplete data are discarded even when they could be valuable and the causal effects of actions on the future are often overlooked. This thesis confronts these limitations by developing four novel deep learning frameworks across two domains: static resource management and dynamic mobility operations. For on-street parking management, this thesis first introduces an attention-based multi-graph model for prediction, which demonstrates that fusing heterogeneous data sources with multiple dynamic graphs can capture more complex spatiotemporal dynamics for higher accuracy. Subsequently, a Hierarchical Deep Reinforcement Learning (HiDRL) framework is presented for real-time assignment, proving to be uniquely scalable to a dynamic number of vehicles and robust to the uncertainties of a mixed vehicle environment. For urban traffic and fleet optimisation, the thesis first proposes an Iterative Multi-task Framework for Completion and Prediction (ImFCP). By synergistically unifying these two tasks, this framework provides more robust traffic forecasts from the incomplete data from real-world sensor networks. Finally, to address a critical “causal blind spot” in conventional fleet management, a Causal-Aware Deep Reinforcement Learning (CADRL) framework is introduced. This model enables an agent to perform counterfactual reasoning about its own impact on passenger demand, allowing it to learn non-myopic and proactive repositioning policies. Validated on large-scale, real-world datasets, each framework demonstrates superior performance over state-of-the-art baselines. Collectively, this research makes a significant contribution by moving beyond simple pattern recognition towards more dynamic and causally aware models, providing a robust foundation for the future of intelligent mobility management.
Fields of Study & Subjects
Traffic Prediction and Management Techniques
Smart Parking Systems Research
Transportation Planning and Optimization
Publication details
- Journal / source
- DR-NTU (Nanyang Technological University)
- Publication date
- 1 January 2026
- Language
- EN
- Document type
- Thesis