Liu, Xin (School of Transportation, Southeast University), HUO, JINBIAO (Southeast University), Zhou, Zhen (Southeast University), Hu, Zhitao (Southeast University), Gu, Ziyuan (Southeast University), Liu, Zhiyuan (Southeast University)

Advancing Simulation-Based Optimization with LLMs for Traffic Problems

Scheduled for presentation during the Invited Session "S18b-Innovative Applications of LLM in Multimodal Transportation Systems" (TH-EA-T18), Thursday, November 20, 2025, 14:50−14:50, Southport 3

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

This paper introduces a novel framework for addressing expensive-to-evaluate simulation-based optimization (SBO) problems in traffic systems by incorporating the global reasoning capabilities of large language models (LLMs). Existing SBO methods commonly rely on random, data-driven search strategies that lack physical interpretability and are often inefficient—particularly during the early stages of optimization. To address this limitation, we design an LLM-assisted Bayesian Optimization (BO) framework in which the LLM acts as a domain expert, providing global search suggestions informed by simulation outputs and problem-specific context.

Specifically, the proposed framework incorporates LLM-inferred knowledge to dynamically define a high-quality sampling region. This region is then explored using conventional BO guided by acquisition functions. We demonstrate the effectiveness of this approach using a traffic signal control problem. Experimental results highlight two key benefits: (1) the LLM-enhanced BO identifies high-quality solutions earlier in the optimization process, and (2) it improves the stability of the optimization by mitigating premature convergence. Overall, this work introduces a generalizable and extensible methodology for embedding domain knowledge into simulation-based optimization through LLMs, offering a new paradigm for traffic system decision-making.