Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

UrbanLLM: Autonomous Urban Activity Planning and Management with Large Language Models

Yue Jiang, Qin Chao, Yile Chen, Xiucheng Li, Shuai Liu, Gao Cong

TL;DR

The paper addresses the need for autonomous handling of complex urban activity planning tasks by coordinating a diverse set of spatio-temporal models. It introduces UrbanLLM, a fine-tuned Llama-2-7B system that decomposes natural-language urban queries into 13 spatio-temporal sub-tasks, matches each to suitable models from a large model zoo, and generates comprehensive responses. Through a learning phase with a self-instruct dataset and a three-stage inference pipeline (spatio-temporal analysis, model matching, results generation), UrbanLLM significantly outperforms strong baselines such as Llama-3 and GPT-4o, and demonstrates robust generalization to other cities. The work reduces human workload in urban planning tasks and offers a scalable, autonomous approach for practical urban decision-support across diverse urban environments.

Abstract

Location-based services play an critical role in improving the quality of our daily lives. Despite the proliferation of numerous specialized AI models within spatio-temporal context of location-based services, these models struggle to autonomously tackle problems regarding complex urban planing and management. To bridge this gap, we introduce UrbanLLM, a fine-tuned large language model (LLM) designed to tackle diverse problems in urban scenarios. UrbanLLM functions as a problem-solver by decomposing urban-related queries into manageable sub-tasks, identifying suitable spatio-temporal AI models for each sub-task, and generating comprehensive responses to the given queries. Our experimental results indicate that UrbanLLM significantly outperforms other established LLMs, such as Llama and the GPT series, in handling problems concerning complex urban activity planning and management. UrbanLLM exhibits considerable potential in enhancing the effectiveness of solving problems in urban scenarios, reducing the workload and reliance for human experts.

UrbanLLM: Autonomous Urban Activity Planning and Management with Large Language Models

TL;DR

The paper addresses the need for autonomous handling of complex urban activity planning tasks by coordinating a diverse set of spatio-temporal models. It introduces UrbanLLM, a fine-tuned Llama-2-7B system that decomposes natural-language urban queries into 13 spatio-temporal sub-tasks, matches each to suitable models from a large model zoo, and generates comprehensive responses. Through a learning phase with a self-instruct dataset and a three-stage inference pipeline (spatio-temporal analysis, model matching, results generation), UrbanLLM significantly outperforms strong baselines such as Llama-3 and GPT-4o, and demonstrates robust generalization to other cities. The work reduces human workload in urban planning tasks and offers a scalable, autonomous approach for practical urban decision-support across diverse urban environments.

Abstract

Location-based services play an critical role in improving the quality of our daily lives. Despite the proliferation of numerous specialized AI models within spatio-temporal context of location-based services, these models struggle to autonomously tackle problems regarding complex urban planing and management. To bridge this gap, we introduce UrbanLLM, a fine-tuned large language model (LLM) designed to tackle diverse problems in urban scenarios. UrbanLLM functions as a problem-solver by decomposing urban-related queries into manageable sub-tasks, identifying suitable spatio-temporal AI models for each sub-task, and generating comprehensive responses to the given queries. Our experimental results indicate that UrbanLLM significantly outperforms other established LLMs, such as Llama and the GPT series, in handling problems concerning complex urban activity planning and management. UrbanLLM exhibits considerable potential in enhancing the effectiveness of solving problems in urban scenarios, reducing the workload and reliance for human experts.
Paper Structure (17 sections, 6 equations, 5 figures, 5 tables)

This paper contains 17 sections, 6 equations, 5 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. UrbanLLM
  4. Urban Activity Planning Learning
  5. Spatio-Temporal Analysis
  6. Model Matching
  7. Results Generation
  8. Experiments and Results Discussion
  9. Experimental Setup
  10. Performance of UrbanLLM
  11. Ablation Study
  12. Visualization and Generalization
  13. Conclusions
  14. Planning and Management Tasks Combinations
  15. Task Understanding Prompt
  16. ...and 2 more sections

Figures (5)

  • Figure 1: The process of solving a real-world problem in urban scenarios involving urban activity planning by human experts.
  • Figure 2: The overall process of the proposed UrbanLLM framework. The urban activity planning learning phase is on the left and the inference phase is on the right.
  • Figure 3: A sample of detailed prompt template used in the learning phase and the process of a new query solved by UrbanLLM in the inference phase.
  • Figure 4: Visualization of responses and results of parking lot occupancy prediction for Marina Square Carpark. GPT-4o needs additional input data highlighted in green for prediction, while UrbanLLM retrieves corresponding data automatically and produces more accurate prediction.
  • Figure 5: Demonstration of generalization ability across cities in UrbanLLM. The upper part is sample user queries in Beijing and New York City, and the lower part presents the resonable outcomes of spatio-temporal task decomposition.