DDRL: A Decentralized Deep Reinforcement Learning Method for Vehicle Repositioning

Online Ride-hailing System improves the efficiency of vehicle utilization and the urban transportation. However, the imbalance between supply and demand is still a problem. To solve this problem and improve resource utilization efficiency, a Decentralized Deep Reinforcement Learning Method (DDRL) for vehicle repositioning is proposed. In DDRL, each vehicle is modeled as an independent agent and dispatched according to its own state to rebalance its local supply and demand. Thus, the global rebalance problem is divided into many small local rebalance problems. First, a new reward evaluation method is proposed and the long-term global reward in traditional reinforcement learning is transformed into many short-term local rewards. Second, a unified algorithm is designed by learning all the decentralized agents' sample data. Finally, the weight matrix of the state is introduced to magnify the differences between the states of adjacent vehicles. Experiments are carried out and the effectiveness of DDRL is verified.