Look-Ahead Insertion Policy for a Shared-Taxi System Based on Reinforcement Learning

This paper proposed a reinforcement learning method to improve the level-of-service (LOS) for a shared-taxi system. In practice, shared-taxi operators usually insert a new arrival request into a vehicle routing system that can minimize current total waiting time and detour distance. However, the LOS of a shared-taxi system does not involve only the total waiting time and detour distance but also the quantity of serviced trip volume. If operators emphasize only on the reduction of the total waiting time and detour distance for current requests, the transport capacity of a shared-taxi system can be excessively expended and cannot reflect future requests effectively. This could lead to a high rejection rate for future requests and damage the global LOS. The proposed reinforcement learning method takes into account the uncertainty of future requests and can make a look-ahead decision to help the operator improve the global LOS of a shared-taxi system. We also tested the proposed method on large-scale networks to verify the performance of the method.