M2MTR: Reposition Idle Taxis in the Many-to-Many Manner with Multi-agent Reinforcement Learning

Ride-hailing apps, such as Didi and Uber, allow people to easily request a ride by inputting their desired origin and destination locations. Due to transportation system complexity and vast city areas, uneven distribution of vehicle supply versus rider demand frequently occurs. This can lead to overcrowded areas with insufficient taxis or sparsely populated zones with abundant empty taxis. To balance supply and demand, many studies have proposed taxi-repositioning methods. However, recent studies limit to repositioning taxis in the one-to-one manner. In this paper, we propose the M2MTR method that can reposition idle taxis in the many-to-many manner. We define the reposition task as a partially observable Markov decision process and define the optimization objectives. To find good reposition strategies, we propose the M2MTR method that is a variation of a multi-agent cooperative A2C method. To make models converge quickly, we design the rewards delicately. To update the policy networks efficiently, we design a local reward combiner. We build an environment simulator to train and evaluate M2MTR. Extensive experiments on real datasets show that M2MTR outperforms other three baseline algorithms. The reposition strategies obtained from M2MTR can make supply and demand more balance, can increase the response rates, and can reduce response time of taxis.