Taxi Dispatch Research Under Different Willingness Of Sharing Based On Reinforcement Learning

With the improvement of living standards of urban residents,demand responsive public transport like taxi has gradually become one of the main choices for people to travel.However,there are some problems in the existing taxi reposition methods,such as the modelling accuracy of the reposition model,the delay of the demand response reposition strategy,the insufficient consideration of the passengers’ willingness of ride-sharing,etc.In view of the above problems,the following research contents are carried out in this paper: Firstly,light GBM,which has a good accuracy on both total and distribution prediction of taxi travel demand,is selected as the taxi demand forecasting algorithm by balancing the predict time and effect.Secondly,based on the hexagonal grid matrix,the predicted demand and the number of dispatchable vehicles in every grid are used as the state;the number of taxis repositioned from each grid to the surrounding grids is used as the action;the average waiting time of passengers and the average dispatching distance of taxis in the dispatching area,combined with the penalties as the reward,a taxi reposition model based on A3 C network is constructed.Thirdly,considering the different passengers’ willingness of ride-sharing,a ride-sharing model,combined the conditions of occurrence of the taxi ride-sharing,the reward function based on the result of ride-sharing,the generation function based on the value of ridesharing,is constructed to optimize the reposition model with ride-sharing behavior,and simulates the ride-sharing situation in the simulation,aiming to analyze the influence of different willingness of ride-sharing in taxi reposition strategy.Finally,based on the above vehicle reposition and ride-sharing model,relying on the actual taxi operation data to build a simulation environment,the proposed reposition strategy is verified,and compared with the on-demand reposition method.Simulation results show that reinforcement learning based reposition strategy can improve the reward score of simulation environment feedback by about 13%;When the demand pressure rises,the proposed dispatching strategy is more effective than on-demand dispatching;When ride-sharing is considered,reinforcement learning and on-demand reposition strategy both achieve a certain optimization effect on the average waiting time of passengers after the increase of passengers’ willingness to share,but it is difficult to improve the corresponding vehicle dispatching distance.However,the improvement effect of the former on the average waiting time is better than that of on-demand strategy,and the higher the willingness to share,the more obvious the optimization effect.Therefore,the reposition strategy proposed in this paper based on reinforcement learning,can optimize the traditional scheduling method according to the verified different situations,and can get more optimization effect when the passenger’s willingness to share increases.This strategy provides a new idea for the management and scheduling of taxi or online car hailing,and has a certain contribution to the follow-up research of taxi scheduling problem.