Explainable AI-based Game-Theoretic Driver Model in Lane-Merge Scenarios = 차선 합류 상황에서의 설명 가능한 AI 기반 게임이론적 운전자 모델

다국어 초록 (Multilingual Abstract)

Human-friendly decision-making has become an essential requirement for autonomous vehicles operating in interactive traffic environments. Although utility-based driver models provide an interpretable framework for autonomous driving decisions, conventional utility functions are typically designed using heuristic assumptions and therefore fail to explicitly reflect the environmental factors that influence human driving decisions.
This paper proposes an Explainable AI-based game-theoretic driver model for lane-merging scenarios by incorporating human decision-making factors into a utility-based decision-making framework. First, Explainable Artificial Intelligence （XAI） is employed to identify the environmental factors governing human driving decisions from real-world driving data. Based on the extracted factors, a human-inspired utility function is developed to quantitatively represent the reasoning process of human drivers. The proposed utility is then integrated into a Stackelberg game-based driver model to generate interactive driving decisions while explicitly considering the responses of surrounding vehicles.
To evaluate the proposed framework, continuous driving trajectories are transformed into discrete longitudinal and lateral driving decisions, enabling direct decision-level comparison between the generated decisions and human driving behavior. Six representative lane-merging scenarios selected from the INTERACTION dataset are used for validation. Qualitative comparisons demonstrate that the proposed driver model more accurately reproduces the decision sequence and lane-changing timing observed in human driving behavior than a conventional utility-based model. Furthermore, quantitative evaluation using sequence-based metrics, including accuracy, precision, recall, F1-score, and segmental edit score, confirms the effectiveness of the proposed utility in representing human driving decisions.
The proposed framework provides a systematic methodology for integrating explainable human decision analysis into utility design and demonstrates that incorporating human decision-making factors improves the ability of utility-based driver models to generate human-friendly decisions. The presented approach establishes an interpretable framework for developing interactive autonomous driving systems capable of generating human-friendly driving behavior.

목차 (Table of Contents)

• List of figures ·····I
• List of tables ····· III
• Abstract ····· IV
• I. Introduction ····· 1
• 1. Motivation ····· 1
• 2. Problem Definition ····· 3
• 3. Objectives ····· 5
• 4. Contributions ····· 6
• II. Literature Review ····· 8
• 1. Human-Friendly Driving Behavior Analysis ····· 9
• 2. Decision-Making Models for Autonomous Driving ····· 12
• 3. Human-Friendly Decision-Making Modeling for Autonomous Driving ····· 15
• III. Extraction of the Decision-Making Factors in Lane-Merge Scenarios ····· 17
• 1. Feature Definition ····· 18
• 1） Dataset and Analysis Scenario ····· 18
• 2） Data Preprocessing ····· 21
• 3） Merging Decision-Making-related Feature Definition ····· 22
• 4） Target Variable Definition ····· 26
• 2. XGBoost-based Merging Maneuver Prediction ····· 27
• 1） XGBoost-based Prediction Model ····· 27
• 2） Model Performance Evaluation ····· 29
• 3. SHAP-based Decision-Making Factors Identification ····· 35
• 1） Global Importance Analysis ····· 36
• 2） Main Effect Analysis of Features ····· 39
• 3） Interaction Effect Analysis between Features ····· 51
• 4） Discussion of Decision-Making Factors ····· 56
• IV. Utility Function Formulation Using Decision-Making Factors ····· 58
• 1. Utility Structure of the Driver Model ····· 59
• 2. Baseline of Safety Utility in Lane-Merge Scenarios ····· 60
• 3. Proposed Safety Utility based on Decision-Making Factors ····· 62
• 1） Proposed Lane-Keeping Utility ····· 62
• 2） Proposed Lane-Merging Utility ····· 68
• V. Stackelberg Game-based Driver Model Design ····· 73
• 1. Interactive Driving Game Framework Design ····· 74
• 2. Stackelberg-based Decision-Making ····· 76
• VI. Validation ····· 78
• 1. Decision-Level Representation of Human Driving Behavior ····· 79
• 2. Decision-Level Validation ····· 82
• VII. Conclusion ····· 88
• Reference ····· 90
• Abstract （Korean） ····· 98