A MARL Framework for Rebalancing Dockless Bike-Sharing Systems

The dockless Bike-Sharing Systems (BSSs) have grown in recent decades, serving as a means of urban mobility capable of reducing traffic congestion and environmental pollution. Compared to conventional docked systems, these systems simplify the process for users, allowing them to pick up and return bicycles anywhere within the service area. However, the lack of physical stations and the asymmetry of demand lead to spatiotemporal imbalances: the accumulation of bicycles at certain points and their complete absence at others severely compromises the system’s effectiveness and reliability. To address this issue, this thesis proposes a solution based on dynamic rebalancing that utilizes Multi-Agent Reinforcement Learning (MARL) techniques. The problem has been formulated as a Markov Decision Process (MDP), in which a fleet of intelligent agents controls the movements and logistical operations (pickup, delivery, recharging) of vehicles tasked with rebalancing demand. To represent the complex topology and operational dynamics of the road network, the simulation environment was modeled using a graph structure, dividing the area into cells. The system state is processed via a graph neural network architecture, which allows agents to extract spatial features and identify critical areas in real time. As a case study, the model was tested and validated on a real-world road network in Manhattan (New York, USA). Numerous simulation tests were conducted to assess the agent’s learning ability and improve its performance, demonstrating the effectiveness of the proposed approach through a reduction in service failures and an even distribution of vehicles across the area.