Research On Taxi Dispatching Method Based On Multi-Agent Reinforcement Learning

With the continuous acceleration of China’s urbanization process and the rapid growth of urban population,it has brought great pressure to the original urban infrastructure,and the contradiction between the demand for transportation and the supply of road facilities has become increasingly prominent.As a very important part of urban operation,the operation efficiency of transportation system not only directly affects people’s travel experience and quality of life,but also has a crucial impact on the economic development of the whole city.Therefore,how to improve the operation efficiency of urban transportation system has become a key issue in urban development.The implementation of public transport services plays a very positive role in relieving the pressure of the transportation system,helping to reduce the number of motor vehicles and reduce road load.As a kind of public transportation,taxi is more and more important in urban transportation because of its convenient and comfortable travel experience and the flexibility to reach any destination.Like other serviceoriented enterprises,taxi companies rely on the income of individual taxi drivers and constantly explore ways to increase profits.But the taxi service is not perfect.Generally,taxi drivers go to some hot areas to find passengers with their long-term business experience.This will cause a large number of taxi resources in a certain area,and the supply exceeds the demand.However,some areas are short of taxi resources,and the supply exceeds the demand,which can not give full play to the carrying capacity of the taxi system.However,the general taxi scheduling method is based on the current realtime demand situation.If the global imbalance between vehicle supply and demand is serious,the time for taxi to arrive at the passenger’s location will increase,resulting in the decline of passenger travel experience.Therefore,this paper designs a taxi scheduling algorithm based on Multi-Agent Reinforcement Learning to schedule taxis from a global perspective,so as to balance the supply and demand of taxis in different regions of the city,improve the coverage of taxi demand in different regions,and maximize the utilization of taxi resources.This paper uses one month’s taxi operation data in the history of Shanghai.Firstly,Hadoop platform and pig script language are used to process the data,remove irrelevant fields and error information,and prepare for the follow-up work.Then,in order to facilitate scheduling,the research area is divided into grids of the same size,and the transfer cost between each pair of grids in different time periods is calculated.Then the number of available vehicles and taxi requests in each time period of each grid is counted from the processed historical information.In order to achieve better and faster supply and demand balance,we hope to know the taxi demand in each grid in the next period,and then schedule vehicles according to the number of taxi requests in each grid in the future.Therefore,this paper uses a variety of time series prediction methods to predict the number of requests in each grid in the future,compares and analyzes these prediction methods,and finally selects the prediction result of the best method as the approximation of the real situation in the future.After getting the number of taxi requests in each grid in the future,we can use the proposed scheduling model based on Multi-Agent Reinforcement Learning to schedule vehicles according to the current vehicle distribution and the number of taxi requests in the future,and finally achieve the balance of taxi supply and demand in each sub region.In the experimental part,the proposed scheduling model is compared with another region based scheduling algorithm,and it is found that the results of this model have achieved good improvement both on weekdays and weekends,30.99% on weekdays and 29.08% on weekends,which has certain guiding significance for taxi operation strategy and scheduling planning.