Rebalancing Docked Bicycle Sharing System with Approximate Dynamic Programming and Reinforcement Learning

The bicycle, an active transportation mode, has received increasing attention as alternative in urban environments worldwide. However, effectively managing the stock levels of rental bicycles at each station is challenging demand vary with time, particularly when users are allowed to return any station. There a need for system-wide management bicycle by transporting available from one another. In this study, rebalancing model based on Markov decision process (MDP) developed using real-time dynamic programming method and reinforcement learning considering system characteristics. pickup demands stochastic continuously changing. As result, proposed framework suggests best operation option every 10 min realized variables future predicted random forest method, minimizing expected unmet demand. Moreover, we adopt custom prioritizing strategies reduce number action candidates operator computational complexity practicality MDP framework. Numerical experiments demonstrate that outperforms existing methods, such short-term static lookahead policies. Among suggested strategies, focusing stations larger error prediction was found be most effective. Additionally, effects various safety buffers were examined.