DeepBike: A Deep Reinforcement Learning Based Model for Large-scale Online Bike Share Rebalancing

Abstract

Abstract Bike share systems (BSSs), as a potentially environment-friendly mobility mode, are being deployed globally. To address spatially and temporally imbalanced bike and dock demands, BSS operators need to redistribute bikes among stations using a fleet of rebalancing vehicles in real-time. However, existing studies mainly generate BSS rebalancing solutions for small-scale BSSs or subsets of BSSs, while deploying small-size rebalancing fleets. How to produce online rebalancing solutions for large-scale BSS with multiple rebalancing vehicles to minimize customer loss is critical for system operation yet remains unsolved. To address this gap, we proposed a deep reinforcement learning based model — DeepBike — that trains deep Q-network (DQN) to learn the optimal strategy for dynamic bike share rebalancing. DeepBike uses real-time states of rebalancing vehicles, stations and predicted demands as inputs to output the long-term quality values of rebalancing actions of each rebalancing vehicle. Rebalancing vehicles could work asynchronously as each individually runs the DQN. We compared the performance of the proposed DeepBike against baseline models for dynamic bike share rebalancing based on historical trip records from Divvy BSS in Chicago, which possesses more than 500 stations and 16 rebalancing vehicles. The evaluation results show that our proposed DeepBike model was able to better reduce customer loss by 111.09% and 57.6% than the mixed integer programming and heuristic-based models, respectively, and increased overall net profits by 101.26% and 220.01%, respectively. The DeepBike model is effective for large-scale dynamic bike share rebalancing problems and has the potential to improve the operation of shared mobility systems.