Prescriptive strategy selection in station-based car sharing: Profit optimization using reinforcement learning

Many car sharing operators struggle to operate at profit. In previous work, it has been shown that trip selection is an important lever to make car sharing systems more profitable and efficient by enabling them to automatically decide whether to accept or reject customer requests. By actively selecting customer requests to be served, car sharing providers are able to implement operative strategies to increase profit, namely increasing collected base fares, increasing collected time-/distant-dependent fares, or reducing operational costs such as relocation costs. However, trip selection has mostly been investigated assuming knowledge of future demands for a fixed horizon and can thus not be unrestrictedly applied in real-time. This paper suggests a prescriptive algorithm, namely a deep reinforcement learning approach (RLA), to solve the trip selection problem solely based on real-time information. Being a machine learning approach, RLA learns which of the above strategies has the highest potential, without the need to specify the booking regime in advance. Based on a simulation of a real car sharing system, we can show that the novel approach can resemble the general structure of optimal offline solutions that were obtained given global information in large parts. It significantly outperforms solutions provided by the state-of-the-art approach widely applied in practice and shows robustness by maintaining the lead even under changing conditions. Narrowly analyzing the solutions found by the novel approach, we apply it to answer complex managerial questions like determining promising relocation rates. • Basic strategies to run car sharing profitably are introduced. • A prescriptive reinforcement learning strategy is applied to solve trip selection. • It detects efficient strategies to increase profit in little computational time. • It significantly outperforms current state-of-the-art technologies. • Provided solutions are structurally closer to optimal offline solutions.