Speaker
Description
Introduction
In the era of smart campus development, emerging technologies such as AI and digital twins have revolutionized educational systems, yet public space management remains reliant on traditional, experience-driven approaches. This gap is evident in critical areas like campus dining spaces, where imbalanced utilization and inefficient resource allocation persist due to insufficient integration of big data and intelligent technologies. Focusing on Tongji University—a high-density urban campus in Shanghai—this study addresses these challenges by developing an agent-based simulation platform to optimize dining space management. Leveraging multi-source data, including 6.3 million campus smart card (CSC) records, geospatial information, and survey responses, we construct a decision-support framework that combines large language models (LLMs) with spatiotemporal behavior analysis to predict user behavior and evaluate management strategies.
Approach and Methodology
The research integrates heterogeneous datasets from May 2023, covering seven campus dining halls, CSC transaction logs (capturing spatiotemporal trajectories at restaurants, classrooms, and gates), class schedules, and individual attributes (e.g., gender, major). A novel LLM-driven agent framework is designed to simulate dining choices, comprising five modules:
- Agent Attributes: Defines individual characteristics (e.g., academic role, schedule).
- Long-Term Memory: Stores campus environmental data and personal dining history.
- Pathfinding Tool: Employs Dijkstra’s algorithm for route optimization.
- Real-Time Context Inputs: Integrates dynamic variables (e.g., updated schedules, crowding alerts).
- Decision Outputs: Generates dining choices and navigational paths.
The model is trained on 20 days of CSC data, validated against a one-day test set and traditional discrete choice models (DCM). Scenarios including staggered class dismissals, real-time crowding information provision, and comprehensive dining quality upgrades (e.g., improved meal variety, service efficiency, and ambient conditions) are simulated to assess spatial utilization efficiency.
Results
The LLM-based agent achieved an 85% accuracy in predicting dining choices, outperforming DCM (76%). Scenario analyses revealed that combining staggered class dismissals with real-time crowding information provision reduced peak-hour congestion by 28% and average wait times by 12 minutes. Meanwhile, comprehensive dining quality upgrades decreased the maximum disparity in space utilization across dining halls from 51% to 27%, promoting equitable spatial distribution. These results demonstrate the framework’s capacity to address both temporal crowding and spatial imbalance.
Discussion and Conclusion
This study advances urban spatial governance by embedding LLMs into agent-based modeling, enabling the simulation of complex human-environment interactions under multidimensional constraints (e.g., schedules, preferences, spatial layouts). Unlike static DCM approaches limited to linear variable relationships, the proposed framework dynamically adapts to contextual changes, incorporates memory-driven learning, and processes heterogeneous data streams—aligning closely with the nonlinear, adaptive nature of human decision-making. By translating granular behavioral predictions into actionable strategies, the platform bridges the gap between data-driven insights and spatial management, fostering resilient, user-centric campus environments.
The methodology offers scalable applications for urban public spaces, including commercial hubs and transit nodes, where balancing efficiency and equity remains a challenge. Future work will integrate real-time IoT data streams and participatory stakeholder feedback to enhance adaptability. This research underscores the transformative potential of AI-augmented decision-support systems in achieving evidence-based, inclusive urban governance—a critical step toward building smarter, more responsive cities.
References
[1]王德 and 蔚丹 (2023) ‘空间行为研究的视角与技术范式’, 城市规划, 47(9), pp. 4–11.
[2]王德 and 胡杨 (2022) ‘城市时空行为规划：概念、框架与展望’, 城市规划学刊, (1), pp. 44–50.
[3]Dang, P. et al. (no date) ‘A large language model-based agent for wayfinding: simulation of spatial perception and memory’, Cartography and Geographic Information Science, 0(0), pp. 1–20.
[4] Zhang, X. et al. (2018) ‘Students performance modeling based on behavior pattern’, Journal of Ambient Intelligence and Humanized Computing, 9(5), pp. 1659–1670.
| Keywords | spatiotemporal behavior; smart campus; decision support systems; agent-based modeling; LLMs |
|---|---|
| Best Congress Paper Award | Yes |
