Spatial-temporal pricing for ride-sourcing platform with reinforcement learning. (September 2021)

• Record Type: Journal Article
• Authors: Chen, Chuqiao; Yao, Fugen; Mo, Dong; Zhu, Jiangtao; Chen, Xiqun (Michael)
• Is Part Of: Transportation research. Volume 130(2021)
• Publication Date: 2021-09
• Subjects: Ride-sourcing -- Spatial–temporal pricing -- Agent-based modeling and simulation -- Reinforcement learning -- Proximal policy optimization (PPO); Transportation -- Periodicals; Transportation -- Technological innovations -- Periodicals; 388.011
• DOI: 10.1016/j.trc.2021.103272
• Languages: English
• ISSNs: 0968-090X

Abstract

Highlights: Develop reinforcement learning enhanced agent-based modeling & simulation framework. Integrate simulation environment with proximal policy optimization algorithm. Build feed-forward critic and actor neural networks to seek the optimal pricing strategy. Tackle spatial-temporal pricing problem for a ride-sourcing platform. Explore the performance of different pricing strategies on a real urban network. Abstract: Ever since the emergence of ride-sourcing services, the spatial–temporal pricing problem has been a hot research topic in both the transportation and management fields. The difficulty lies in simultaneously obtaining the optimal multivariable solution for spatial pricing and sequential solution for dynamic pricing, considering the heterogeneity, dynamics, and imbalance of on-demand ride supply/demand. Due to this problem's complexity, most studies have simplified the modeling setting and omitted the complicated matching and waiting process between drivers and passengers. To go beyond the existing models, this paper proposes a reinforcement learning enhanced agent-based modeling and simulation (RL-ABMS) system to reveal the complex mechanism in the ride-sourcing system and tackle the problem of spatial–temporal pricing for a ride-sourcing platform. The reinforcement learning approach proximal policy optimization (PPO) is implemented in the RL-ABMS system, where two feed-forward neural networks are built as critic and actor. The critic judges the goodness of the current state, and the actor generates the optimal pricing strategy. Compared with the fixed pricing strategy, the experimental results on a real-world urban network show that dynamic pricing raises the platform's profit to 1.25 times, and spatial–temporal pricing even raises it to 1.85 times. Besides, the number of idle drivers/vehicles has significantly dropped under the spatial–temporal pricing strategy, which indicates that our proposed strategy has a remarkable effect on coordinating supply and demand in the ride-sourcing market.