Deep Reinforcement Learning Optimizes Dockless Bike-Sharing Rebalancing

A research article introduces a dynamic Deep Reinforcement Learning (DRL) approach for optimizing dockless bike-sharing systems, overcoming the challenges posed by periodic, system-wide rebalancing efforts. The research utilizes a graph-based simulator to model the service and formulates rebalancing as a Markov decision process. In this framework, a DRL agent manages a single truck in real time, performing localized pick-ups, drop-offs, and charging based on spatiotemporal criticality scores. Real-world data experiments reveal substantial decreases in availability failures with a minimal fleet size, while also addressing spatial inequality and mobility deserts. This method highlights the effectiveness of learning-based rebalancing for enhancing shared micromobility. The findings are available on arXiv in the Electrical Engineering and Systems Science section.