이용자 선호 인지형 (자율주행) DRT 배차에 관한 연구 : 불확실한 호출수요를 고려하는 문맥적 밴딧 접근법- = A Study on Preference-Aware (Autonomous) DRT Dispatching - A Contextual Bandit Approach Consid ering Uncertain Demand

다국어 초록 (Multilingual Abstract)

Demand-Responsive Transit (DRT) systems offer flexible mobility in areas where fixed-route transit is impractical, yet most operational systems rely on system-centric dispatching that overlooks passenger experience. This thesis proposes a Preference-Aware Adaptive Dispatch Framework that incorporates stated-preference (SP) behavioral values into real- time vehicle assignment. The framework combines a traditional feasibility-based insertion stage with a second stage that re-ranks feasible assignments using contextual bandit algorithms—LinUCB and Thompson Sampling—initialized with SP-derived utility coefficients capturing sensitivity to waiting time, walking distance, and detour. The system is implemented within a high-fidelity simulation of the Hwaseong Living-Lab environment. Feasible vehicleroute insertions are generated under strict operational constraints including capacity, detour limits, service-level windows, and walking thresholds. These candidates are then evaluated using a passenger-utility model that enables the dispatcher to learn and adapt over time. The use of SP priors eliminates the cold-start problem and ensures behavioral alignment from the first day of operation. Across 900 scenario-days covering five demand patterns and four uncertainty regimes, the preference-aware policies consistently outperform a standard greedy baseline. Thompson-SP reduces mean waiting times by 30–40%, increases service rates by 20–31%, and improves high-percentile reliability while maintaining full feasibility. The approach also enhances service equity, reducing spatial disparities in waiting times and improving outcomes for wait- sensitive and walk-averse passenger segments. This research demonstrates that integrating behavioral preferences with online learning produces substantial improvements in DRT performance and fairness without increasing fleet size. The proposed framework offers a practical foundation for user-centered, adaptive, and equitable dispatching in next-generation DRT systems.

목차 (Table of Contents)

• I. INTRODUCTION 1
• 1.1 Background and Motivation 1
• 1.2 Problem Statement 2
• 1.3 Research Objectives 3
• 1.4 Research Questions and Hypotheses 4
• 1.5 Contributions of This Thesis 5
• 1.6 Significance and Practical Impact 5
• 1.7 Thesis Organization 6
• II. LITERATURE REVIEW 7
• 2.1 Demand-Responsive Transit Operations and Dispatch Optimization 7
• 2.2 Passenger Behavior, Discrete Choice Modeling, and Stated Preferences 9
• 2.2.1 Foundations of Travel Behavior Modeling 9
• 2.2.2 Empirical Evidence on Waiting and Walking Disutility 9
• 2.2.3 Applications of SP/RP Data in Transit Planning 9
• 2.2.4 Heterogeneity in Passenger Preferences 9
• 2.2.5 Gap 2: Limited Integration of Behavioral Preferences in Real-Time Dispatch 10
• 2.3 Contextual Bandits and Online Adaptive Decision-Making 10
• 2.3.1 Multi-Armed Bandits and the Exploration–Exploitation Trade-off 10
• 2.3.2 Contextual Bandits 10
• 2.3.3 Bandits in Transportation Applications 11
• 2.3.4 Why Bandits Fit the DRT Dispatch Context. 11
• 2.3.5 Gap 3: Lack of Bandit-Based DRT Frameworks 11
• 2.4 Equity and Distributional Impacts in Flexible Mobility 11
• 2.4.1 Foundations of Transport Justice 11
• 2.4.2 Equity Concerns in DRT Systems 12
• 2.4.3 Equity Measurement and Evaluation 12
• 2.4.4 Gap 4: Equity Rarely Considered in Dispatch Algorithms 12
• 2.5 Summary of Literature Gaps and Research Positioning 12
• III. STUDY AREA, DATA, AND STATED-PREFERENCE SURVEY 14
• 3.1 Overview 14
• 3.2 Study Area: Hwaseong Living Lab 14
• 3.2.1 Regional and Administrative Context 14
• 3.2.2 Population, Urban Structure, and Mobility Characteristics 14
• 3.2.3 Transportation Patterns and DRT’s Role 15
• 3.2.4 Geographic Characteristics 16
• 3.2.5 Zoning using H3 hexagonal indexing 16
• 3.3 Network Data and Virtual Stops 17
• 3.3.1 Drive network 17
• 3.3.2 Walk network 18
• 3.3.3 Virtual stop placement 19
• 3.4 Demand Data 21
• 3.4.1 Real-time demand scenarios 22
• 3.4.2 Spatial structure of demand 22
• 3.4.3 Temporal distribution 23
• 3.5 Stated-Preference (SP) Survey 23
• 3.5.1 Survey design 23
• 3.5.2 Attributes and efficient design 23
• 3.5.3 Model estimation 24
• 3.5.4 Utilization of SP Results in the Dispatch Framework 25
• 3.6 Demand Uncertainty Modeling 26
• 3.6.1 Rationale 26
• 3.6.2 Two dimensions of uncertainty 26
• 3.6.3 Technical method: Poisson thinning and augmentation 27
• 3.6.4 Validation 28
• 3.7 Summary 29
• IV. METHODOLOGY 30
• 4.1 Introduction 30
• 4.2 Overview of the Dispatch Framework 30
• 4.3 Stage 1: Feasible Vehicle–Route Insertion Generation 32
• 4.3.1 Proximity Filtering 32
• 4.3.2 Insertion Enumeration 32
• 4.3.3 Feasibility Constraints 32
• 4.3.4 Slate Construction and Top-K Truncation 34
• 4.4 Stage 2: Preference-Aware Adaptive Assignment 34
• 4.4.1 Contextual Bandit Formulation 34
• 4.4.2 LinUCB-SP 35
• 4.4.3 Thompson-SP 35
• 4.4.4 Algorithm Hyperparameters 36
• 4.4.5 Decision Rule and Policies 37
• 4.5 Feature Engineering and Context Representation 37
• 4.6 Reward Function and Utility Modeling 38
• 4.6.1 SP-Based Utility and Reward Specification 38
• 4.6.2 New vs Existing Passengers 39
• 4.7 Initialization and Updating of Segment-Specific Preferences Using Stated-Preference (SP) Priors 40
• 4.7.1 Mapping SP Coefficients to Prior Parameters 40
• 4.7.2 Embedding Logit Segments into the Bandit Model 40
• 4.7.3 How the Three Segments Are Updated During Learning 41
• 4.7.4 Role of Global and Segment-Specific Priors 42
• 4.7.5 Benefits of SP-Based Initialization and Segment-Aware Updating 42
• 4.8 Simulation Engine and Learning Procedure 43
• 4.8.1 Event-Driven Architecture 43
• 4.8.2 Learning Loop 43
• 4.8.3 Experimental Structure 43
• 4.9 Evaluation Metrics 44
• 4.9.1 Passenger Experience Metrics 44
• 4.9.2 Operational Metrics 44
• 4.9.3 Equity and Robustness Metrics 44
• 4.10 Summary 45
• V. RESULTS 46
• 5.1 Introduction 46
• 5.2 Aggregate Performance Across All Conditions 46
• 5.2.1 Improvements in Passenger Experience 46
• 5.2.2 Service Rate Improvements 47
• 5.2.3 Minimal Trade-offs 47
• 5.3 Scenario-Level Performance Differences 47
• 5.3.1 Base Scenario 48
• 5.3.2 Temporal Peaks (S1) 48
• 5.3.3 Spatial Concentration (S2) 48
• 5.3.4 Joint Temporal–Spatial Peaks (S3) 48
• 5.3.5 Smart-Card Derived (S4) 48
• 5.4 Robustness to Demand Uncertainty 49
• 5.4.1 High-Under Scenario (Severe Underestimation) 49
• 5.4.2 High-Over Scenario (1.5× Overestimation) 50
• 5.4.3 Medium and Low Uncertainty 50
• 5.5 Learning Dynamics Across Multiple Days 50
• 5.5.1 Day 1: Cold-Start With SP Priors 50
• 5.5.2 Days 2–5: Convergence to Stable Policies 51
• 5.6 Passenger Equity Analysis 51
• 5.6.1 Equity Across Passenger Segments 51
• 5.6.2 Spatial Equity Across Zone Groups 52
• 5.7 Operational Efficiency and SLA Compliance 53
• 5.7.1 SLA Performance 53
• 5.7.2 Fleet Utilization and Travel Metrics 53
• 5.8 Summary of Findings 54
• VI. DISCUSSION 55
• 6.1 Overview 55
• 6.2 Improvements in User Experience 55
• 6.2.1 Reductions in Waiting Time and Tail Delays 55
• 6.2.2 Service Rate as a Second-Order User Experience Benefit 56
• 6.3 Mechanisms Behind Adaptive Learning and Cold-Start Benefits 56
• 6.3.1 The Role of SP Priors in Mitigating Cold-Start Limitations 56
• 6.3.2 Distinct Learning Dynamics of LinUCB-SP and Thompson-SP 57
• 6.4 Robustness under Demand Uncertainty 57
• 6.4.1 High-Under Scenario: The Most Stringent Test 57
• 6.4.2 High-Over Scenario: Benefits in Oversaturated Conditions 58
• 6.4.3 Reliability Across Medium and Low Uncertainty 58
• 6.5 Equity Outcomes Across Passenger Groups and Spatial Regions 58
• 6.5.1 Equity Across Behavioral Segments 58
• 6.5.2 Spatial Equity Across Zones 59
• 6.6 Operational Feasibility and Efficiency Trade-offs 59
• 6.6.1 SLA Compliance 59
• 6.6.2 Fleet Utilization and VKT 59
• 6.7 Contributions to DRT System Design and Planning 60
• 6.7.1 Practical Deployability in Korean Municipalities 60
• 6.7.2 Complementarity with Existing Routing Engines 60
• 6.8 Limitations and Considerations for Generalization 60
• 6.9 Synthesis and Theoretical Contributions 61
• 6.10 Concluding Remarks 61
• VII. CONCLUSION 62
• 7.1 Overview 62
• 7.2 Key Contributions 62
• 7.3 Implications for DRT System Design and Policy 63
• 7.4 Limitations 63
• 7.5 Directions for Future Research 64
• 7.6 Closing Remarks 65