Dynamic Matching Optimization in Ridesharing System Based on Reinforcement Learning

Modern urban transportation, has concurrently posed environmental challenges such as traffic congestion and increased greenhouse gas emissions. In response to these issues, ridesharing systems have emerged as a viable solution. By fostering ridesharing among individuals with similar travel routes, ridesharing, effectively, optimizes vehicle utilization, offering a sustainable and practical alternative to address contemporary transportation challenges. In this work, we delve into intricacies of dynamic ridesharing systems. Focusing on the dynamic matching problem within ridesharing, we propose a solution leveraging reinforcement learning. Our contribution involves the distinct modeling of two scenarios: one-to-one and one-to-many ridesharing. In the one-to-one scenario, spatiotemporal constraints are considered with the objective of minimizing passengers' waiting times. In the more complex one-to-many scenario, additional constraints are introduced focusing on both minimizing passengers' waiting times and drivers' detour times. The proposed modeling is time-focused assuming that time is a cutting parameter in the decision-making. The results obtained through our experiments demonstrate the system's effectiveness, robustness and adaptability to diverse constraints.