Developing a supervised learning-based simulation method as a decision support tool for rebalancing problems in bike-sharing systems

Bike-Sharing Systems (BSSs) have exploded in popularity worldwide because of their beneficial impacts on traffic, pollution levels, and public health, which has resulted in moving toward a green city. The rebalancing problem, as one of the most important operational problems of such systems, deals with planning bike distribution at different stations. Regarding conducted studies, simulation models are the most common tool for analyzing BSSs and decision-making. This popularity is based on simulation's capabilities in modeling complexities of systems and uncertainty. Despite their advantages, lack of quickness is a significant drawback of simulation-based methods, making them inefficient for real-time decision-making processes, especially in largescale and complex systems. In this regard, this paper introduces a Supervised Learning-Based Simulation (SLBS) method as an alternative to the conventional simulation-based methods dealing with rebalancing problems. SLBS is a huge step toward developing a real-time Decision Support System (DSS) for BSSs. For developing SLBS, firstly, we have developed a simulation model based on real-world assumptions of station-based BSSs and big data analysis of CitiBike, a well-known BSS located in New York City. The simulation model is able to calculate the number of unsatisfied demands (either number of failed pick-ups (FPs) or failed drop-offs (FDs)) as a result of different rebalancing plans. Then, the developed simulation model was used to generate quality and quantity training datasets to train Machine Learning (ML) algorithms involved in SLBS. While these ML models are trained once, SLBS will be capable of predicting the number of unsatisfied demands without running highly time-intensive simulations replications. The results obtained from a wide range of conducted experiments indicate that SLBS, up to 300 times faster than simulation models, can provide predictions with over 90% of R2 Score.