Large Language Model Agent for Mobility Reasoning and Synthesis – MobAgent

The proliferation of time-stamped and geo-tagged human mobility data from sources has immense potential for applications in transportation planning, travel behavior analysis, disease monitoring, and intelligent urban management. However, releasing large-scale mobility datasets for public use is hindered by significant challenges: individual mobility data are highly sensitive due to privacy concerns, and data from single sources are often partial, sparse, noisy, and independent of other datasets. These issues impede the ability to generate complete trip chains required for agent-based simulations and other applications. To address these complex challenges, the MobAgent project aims to develop advanced machine learning methodologies for realistic human mobility synthesis, particularly using Large Language Model (LLM)-agent for mobility reasoning and synthesis. Our research focuses on the following objectives. 1) Develop Deep Generative Models (DGMs) for Realistic Single-Source Mobility Data Synthesis while accounting for data sparsity, noise, and censorship. This involves both modeling Data Imperfections within the generative process, and quantitative evaluation. 2) Enhance DGMs with Attention-Based Mechanisms for Complex Spatiotemporal Dependencies to capture long-range temporal dependencies and spatial correlations across human mobility data and also address Data Heterogeneity. 3) Design a Multimodal LLM-Based Agent Framework for Multi-Source Data Fusion. We will develop a framework to fuse this multimodal data into complete trip chains by utilizing LLMs' capabilities to process various data formats and understand human activity patterns to create specialized agents that interact to process, refine, and integrate synthesized data, ensuring the generation of realistic and coherent mobility sequences. The developed MobAgent will be evaluated using real-world agent-based simulation models in both French and Canadian cities.