关键词: 智能交通, 客货联运, 多智能体强化学习, 需求响应公交, 动态调度, 时空灵活性

Abstract: Demand- responsive transit, a representative mode of a shared mobility system, faces the challenge of efficiently handling travel demand and low vehicle utilization during peak hours. Previous research investigates the optimization of operational scheduling based on the immobile travel demand in the spatiotemporal dimension and has not yet explored the potential improvements in terms of human flexibility. The development of e-commerce has resulted in the rapid growth of freight quantity, but it also faces the problems of truck restrictions in the urban area and low time efficiency of parcel delivery. This paper focuses on designing a new mode of the shared mobility system, and proposes a demand-responsive transit service mode that accommodates both passenger and freight transportation and considers passengers' spatiotemporal mobility. First, the study introduces a new reservation management mechanism for passengers' spatiotemporal flexibility, and formulates the dynamic scheduling process of demand- responsive passenger and freight combined transportation as a Markov decision process in a rolling horizon framework. Each vehicle is considered as an agent, while the information of stops, requests, and vehicles in the environment is integrated as a state of reinforcement learning. The action space is designed to meet the constraints of time, capacity and spatiotemporal mobility adjustment. The reward function of the adaptation model is developed according to the dynamic scheduling characteristics. The method minimizes the total system cost and solves the Markovian decision process model with the Qtran_alt framework based on multi-agent reinforcement learning algorithm. A test was performed in the southern area of Huangpu District, Guangzhou City, and six performance indicators were designed to validate the overall performance of the model and algorithm. The performance was compared between the Qtran_alt and three other multi- agent reinforcement learning algorithms. And the sensitivity analysis was performed for spatiotemporal mobility, time slice interval, overtime penalty factor and freight- to- passenger ratio. The computational results demonstrate that our model can serve 50% of additional freight requests without an increase in total system cost of more than 5.37%, and reach a request service rate of 95.19% with strong universal applicability.

Key words: intelligent transportation, passenger and freight combined transportation, multi-agent reinforcement learning, demand-responsive transit, dynamic scheduling, spatiotemporal flexibility