Title Page
Abstract
Contents
Chapter 1. Introduction 12
1.1. Background 12
1.1.1. Social Background 12
1.1.2. Academic Background 15
1.2. Research Purpose 18
Chapter 2. Literature Review 20
2.1. Concepts 20
2.1.1. Characteristics of Ride-hailing Service 20
2.1.2. Definition of Dynamic Pricing 21
2.1.3. Ride-hailing Service as a Sequential Decision Problem 22
2.2. Dynamic Pricing 23
2.3. Negative Externalities 25
2.4. Implications 28
Chapter 3. Methodology 30
3.1. Contextual Bandit 30
3.2. Temporal Difference Learning 34
3.3. Model Formulation 37
3.3.1. Semi-MDP (Markov Decision Problem) 37
3.3.2. Bootstrapping Structure 40
Chapter 4. Data Description 44
4.1. Data Tables 44
4.2. Descriptive statistics 49
4.3. Spatial Distribution 50
Chapter 5. Results 54
5.1. Base Model 54
5.2. Pickup Waiting Time Reducing Model 63
5.3. Effect of Changes in Price Sensitivity 67
5.4. Effect of Changes in Value Term Coefficient 72
5.5. Applicability of the Model 80
5.5.1. Upper Limit of Surge Multiplier 80
5.5.2. Applicability to Off-peak Hours 80
5.5.3. Formulate Policies to Reduce Pickup Cost 83
Chapter 6. Conclusion 84
6.1. Summary 84
6.2. Limitations and Future Research 85
References 86
국문초록 93
Table 1. Revenue and forecasts of ride-hailing and taxi in Singapore 13
Table 2. Comparison of matching time and pickup waiting time in historical data 17
Table 3. Elements of reward function 39
Table 4. Number of requests by type after preprocessing 46
Table 5. Columns of ride data table 47
Table 6. Columns of ping data table 48
Table 7. Descriptive statistics for Monday, November 16, 2020 49
Table 8. Comparing two terms of trial simulation 55
Table 9. Matching rate and error rate by travel distance 58
Table 10. Matching rate and error rate by time 59
Table 11. Distribution of surge multipliers (Base model) 60
Table 12. Value term and matching rate in base model 62
Table 13. Comparison of base model and pickup cost reducing model 65
Table 14. Comparison to TADA's pricing 66
Table 15. Matching rate according to travel distance with changing price sensitivity 68
Table 16. Matching rate according to time with changing price sensitivity 69
Table 17. Revenue and reward with changing price sensitivity 70
Table 18. Convergence of algorithm with changing price sensitivity 71
Table 19. Matching rate according to travel distance with changing value term coefficient 73
Table 20. Matching rate according to travel distance with changing value term coefficient 75
Table 21. Matching rate according to time with changing value term coefficient 76
Table 22. Revenue and reward with changing price sensitivity 77
Table 23. Convergence of algorithm with changing value term coefficient 79
Table 24. Matching rate by travel distance (off-peak) 81
Table 25. Comparison of base model and pickup cost reducing model (Off-peak hours) 82
Figure 1. Ride-hailing companies 13
Figure 2. Time intervals that constitute a single request 16
Figure 3. Structure of reinforcement learning 30
Figure 4. Difference between a full reinforcement learning problem and a contextual bandit 31
Figure 5. Update of value function 35
Figure 6. Bootstrapping structure of algorithm 41
Figure 7. An example illustrating that updating the algorithm results in new situations 42
Figure 8. Hourly hailing requests of November 16, 2020 44
Figure 9. Spatial scope of study 45
Figure 10. Spatial distribution of origins (left) and destination (right) 51
Figure 11. Spatial Distribution of Origins of Matched (Left) and Unmatched (Right) Requests 52
Figure 12. Spatial Distribution of Destinations of Matched (Left) and Unmatched (Right) Requests 52
Figure 13. The proportion of each surge multiplier converges as iteration progresses 56
Figure 14. The percentage of requests with altered surge multipliers in the subsequent iteration 57
Figure 15. Matching rate by travel distance 58
Figure 16. Matching rate by time 59
Figure 17. Spatial distribution of origins and requests with surge multipliers less than 1 61
Figure 18. Matching rate according to value term 62
Figure 19. Comparison of surge multiplier distribution with base model (Left) 64
Figure 20. Comparison of surge multiplier distribution with base model (Left) 74
Figure 21. Temporal matching rate of off-peak 81
Figure 22. Surge multiplier gap between base model and pickup cost reducing model 83