Spatio-temporal task pricing for shared electric micro-mobility battery-swapping platform with reinforcement learning

Spatial crowdsourcing has emerged in shared electric micro-mobility platforms, compensating occasional drivers (ODs) per task of swapping micro-mobility batteries. As ODs autonomously select tasks only when satisfied with predetermined compensation and travel distance, a traditional uniform pricing strategy results in possible low task completion. To resolve the imbalance between ODs and tasks, this study introduces a spatio-temporal pricing strategy where task prices differ by region and time interval. Considering the daily variations in task distribution and OD availability, the goal is to minimise the platform costs equal to the sum of total OD wages and penalties for uncompleted tasks. The reinforcement learning approach with proximal policy optimisation (PPO) is implemented to generate real-time continuous task prices. A domain-specific masking technique is incorporated to improve the learning process by disregarding the data from inactive grids in loss calculations. Computational results show that the PPO agent strategically raises prices in regions with insufficient ODs according to the OD density level. Further comparison with the mixed integer programming model with perfect information on ODs' willingness-to-accept parameters demonstrates the superior capability of our algorithm in navigating the uncertainties of OD task acceptance. A sensitivity analysis provides insights into the decision of system parameters.