引力视角下出租车调度的增强学习模型

Reinforcement Learning Model of Taxi Dispatching under Gravity Perspective

摘要: 以互联网预约式叫车服务为背景，针对出租车运营中司乘匹配率低、乘客等车时间长和车辆空驶距离长的问题，通过类比万有引力模型，提出了空驶出租车巡游路线调度优化的增强学习控制方法。分别以备选目的地等车乘客数和当前位置空驶出租车数为对应位置的“质量”，而以两地间最短路径长度为两点间“距离”，定义两点间“引力”。通过引力大小，决策空驶车辆的目的地。通过引入增强学习理论，实现空驶出租车巡游路线决策的时空全局最优性，从而避免仅凭当前引力值大小决策路线的短视性。数值算例分析表明：新的调度方法不仅可以提高司乘匹配率和总的运营收入，而且可以减少总的空驶距离。

Abstract: In the setting of internet appointing taxi service, by analogizing the universal gravitation model, a control method of optimizing the cruising routes of vacant taxis was proposed to solve the problems of the low matching rate between taxis and passengers, the long average waiting time of passengers and the long travelling distance of vacant taxis. Using the number of waiting passengers as the mass of the corresponding alternative destination, the number of vacant taxis as the mass of current position, and the length of the shortest path between these two positions as their distance, the gravitation between these two positions can be defined. The next destination of a vacant taxi can be determined using the defined gravitation. By introducing the reinforcement learning theory, the decision of choosing the routes of vacant taxis possesses the global optimality. The short sight of deter-mining the routes only by the size of gravitation can also be avoided. The numerical example shows that the new taxi dispatching method can not only increase the matching rate between taxis and passengers and the total operation income, but also decrease the total travelling distance of vacant taxis.