Using Deep Reinforcement Learning to Coordinate Multi-Modal Journey Planning with Limited Transportation Capacity

Abstract

Multi-modal journey planning for large numbers of simultaneous travellers is a challenging problem, particularly in the presence of limited transportation capacity. Fundamental trade-offs exist between balancing the goals and preferences of each traveller and the optimization of the use of available capacity. Addressing these trade-offs requires careful coordination of travellers’ individual plans. This paper assesses the viability of Deep Reinforcement Learning (DRL) applied to simulated mobility as a means of learning coordinated plans. Specifically, the paper addresses the problem of travel to large-scale events, such as concerts and sports events, where all attendees have as their goal to arrive on time. Multi-agent DRL is used to learn coordinated plans aimed at maximizing just-in-time arrival while taking into account the limited capacity of the infrastructure. Generated plans take account of different transportation modes’ availability and requirements (e.g., parking) as well as constraints such as attendees’ ownership of vehicles. The results are compared with those of a naive decision-making algorithm based on estimated travel time. The results show that the learned plans make intuitive use of the available modes and improve average travel time and lateness, supporting the use of DRL in association with a microscopic mobility simulator for journey planning.