LLM-Based Contextual Understanding of Real-World Driving Behavior and Realistic Reproduction in Traffic Simulation
- 주제(키워드) Driving Behavior , Traffic Flow , Large Language Models , Traffic Simulation
- 발행기관 아주대학교 일반대학원
- 지도교수 Jaehyun So
- 발행년도 2025
- 학위수여년월 2025. 8
- 학위명 석사
- 학과 및 전공 일반대학원 D.N.A.플러스융합학과
- 실제URI http://www.dcollection.net/handler/ajou/000000035292
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Traditional rule based traffic reproduction models often fail to replicate the nuanced and heterogeneous decision making behaviors of human drivers. This research proposes a novel trajectory generation framework based on Large Language Models (LLMs), designed to emulate human like driving decisions through contextual reasoning and behavior reproduction. Therefore, utilizing the Next Generation Simulation (NGSIM) US-101 dataset, we first apply a hybrid unsupervised clustering method that combines Self Organizing Maps (SOM) and K-means++ to classify drivers into six behavioral clusters. To generate context aware driving trajectories, we construct a Chain of Thought (CoT) prompting framework that provides the LLM with ego vehicle state information as well as dynamic interaction data from surrounding vehicles (e.g., velocity, acceleration, headway) via a shared memory mechanism. This setup enables the LLM to simulate the perception judgment action loop typically exhibited by human drivers, and to produce natural language explanations of its driving decisions. The experimental evaluation focuses on a 30 second segment of the NGSIM dataset, comparing the LLM generated trajectories with those produced by IDM based SUMO simulations. Results show that the LLM achieves lower Mean Absolute Errors (MAE) across velocity, acceleration, and headway metrics for most driver clusters. Furthermore, it exhibits more realistic responses in high risk situations (TTC < 1.5s) and better reproduces lane change behaviors. Primarily, the LLM also generates interpretable reasoning for each decision, explicitly indicating whether safety or mobility was prioritized. These findings demonstrate that LLMs can function not merely as trajectory generators but as high level cognitive agents capable of replicating the contextual judgment processes of human drivers. The proposed approach enhances both the behavioral realism and interpretability of traffic simulation, offering new possibilities for autonomous vehicle testing, behavioral policy evaluation, and the future of explainable AI in transportation systems.
more목차
Chapter 1. INTRODUCTION 1
1. Research Backgrounds 1
2. Research Goal and Objectives 4
Chapter 2. LITERATURE REVIEW 7
1. Rule Based Traffic Flow Reproduction Models 7
2. Traffic Flow Reproduction with Driver Characteristics 9
3. Data Driven and AI Based Traffic Flow Reproduction Models 11
4. LLM Based Contextual Modeling of Vehicle Behavior 13
5. Lessons Learned 15
Chapter 3. METHODS 18
1. Data Collection and Preprocessing 19
2. Behavior Clustering with SOM and K-means++ 22
3. Contextual Trajectory Reproduction Using LLM 26
4. Traffic Simulation Modeling 31
5. Evaluation Framework and Performance Analysis Criteria 34
Chapter 4. RESULTS 39
1. Reproduction Accuracy Analysis 39
2. Analyzing Safety and Mobility 45
3. Evaluate the Interpretability of the Model 47
Chapter 5. CONCLUSION 49
1. Conclusions 49
2. Limitations and Future Research 51
REFERENCES 53
초록 60
