E-Bike Station Management using Deep Reinforcement Learning

Bike sharing systems (BSS) are becoming important with the growth of eco-friendly transportation systems. In this regard, there have been developments on predicting the demand for bikes, repositioning of the bikes across the stations and minimizing the imbalance between the supply and the demand across the stations. This paper explores the station management problem and the potential advantages it can have in the overall development of the BSS. Further, we consider the case where the bikes are electric. Thus, unlike the existing works, it introduces additional responsibilities for the stations under the BSS. Unlike the existing works, we propose a new e-bike station management problem which does not require the information about the actions of other stations or the decisions of the trucks which transfer bikes across the stations. The objective of the proposed problem is to maximize the reward for supplying the e-bikes with higher charge level and satisfying the demand for the e-bikes. We formulate the problem as a Markov decision process which is approximated and solved using the deep deterministic policy gradient algorithm. We perform extensive simulations and provide detailed analysis of the station’s actions and average episodic reward. We compare the results of the proposed method with a greedy approach and a method which takes random actions.