Ride-sourcing services have experienced a fast-growing period in this era around the world. The prosperity of the novel mobility mode enriches the current transportation mode, enhances the connection across the city, and quickly becomes one of major travelling choices.

Different to traditional taxi industry, the services for ride-sourcing are mainly provided by some centralized platform, rather than directly by individual drivers. On the one hand, the information of supply and demand sides in the market can be collected and employed with higher completeness and efficiency. On the other hand, such management approaches also generate larger operation challenges, since a larger number of user data have to be properly and simultaneously treated. In the past, ride-sourcing platforms mainly o...[ Read more ]

Ride-sourcing services have experienced a fast-growing period in this era around the world. The prosperity of the novel mobility mode enriches the current transportation mode, enhances the connection across the city, and quickly becomes one of major travelling choices.

Different to traditional taxi industry, the services for ride-sourcing are mainly provided by some centralized platform, rather than directly by individual drivers. On the one hand, the information of supply and demand sides in the market can be collected and employed with higher completeness and efficiency. On the other hand, such management approaches also generate larger operation challenges, since a larger number of user data have to be properly and simultaneously treated. In the past, ride-sourcing platforms mainly optimize their operation decisions based on current or historical market information. However, any current decisions actually have an impact on the future market status for ride-sourcing services, making them possibly not optimal from a long-term perspective. To handle these challenges, an important and applicable idea is to extract future information from the historical data, and utilize the combined current and future information to improve the farsightedness of the operation strategies. Thanks to the wide application of GPS and APP on the smart phones, the data source for ride-sourcing services has been sufficiently fruitful to support the implementation of the idea.

More specifically, the big data for ride-sourcing services can be employed from two different phases. First, the historical data can be utilized to estimate future data of market status, which we call prediction. Second, the historical and estimated future information can be combined to optimize the platform operations, sharing the same spirit with reinforcement learning technology (an important branches of machine learning). However, the amount of data is so large and the property of data is so dynamical that it is difficult for human to directly understand and utilize them in an efficient way. As a result, we have to heavily depend on machine to learn and employ data, which we call machine learning.

Although machine learning has been widely applied in multiple fields, including computer science, robotics, finances, etc., the integration of it into ride-sourcing industry still face great challenges. For prediction phase, the data source and prediction objective can be diversified, while the prediction results are often expected to be generated in an efficient and simultaneous way. Moreover, the ride-sourcing data is widely spread over the whole spaces of the cities, leading to complex intrinsic spatial relationship within different data points. For operational phase, major challenges arise from the proper design of operation strategies with future information available, the possible interaction with other transportation modes, and the development of simulation environments for the iterative improvement of the designed strategies.

To handle these issues, we first design a neural network based prediction model for ride-sourcing services, which can handle multiple important prediction tasks within one framework. Furthermore, the complex and hierachical spatial correlations within the data are discussed and explored in the second study, where another prediction framework is proposed to capture both micro and macro spatial correlations within the traffic data. For operational phase, we first develop a comprehensive and multi-functional simulation platform as the general test bed for all the algorithms training and testing for ride-sourcing services. The platform is constructed on real transportation network, can simulate different behaviors of stakeholders in the market, and have been completely open-sourced. Afterwards, we concentrate on the operation strategy design for ride-sourcing services, where a reinforcement learning (RL) framework is proposed to optimize the core operation of the service, with coordination within different transportation modes. To summarize, we focus on the utilization of machine learning technology to reshape the strategy design logic for the operations in ride-sourcing services, with the aim at higher efficiency, effectiveness, and farsightedness.