A deep reinforcement learning approach for real-time bus operation control using departure timetabling, stop-skipping, and re-routing

Abstract

Bus operation control is an effective way to enhance bus line operational efficiency and provide high-quality intra-city transit services. Traditional bus operation control approaches predominantly focus on static control, while the dynamic nature of transit ridership necessitates real-time control approaches. This study proposes a novel multi-agent Deep Reinforcement Learning (DRL) approach to address the real-time bus operation control problem, using a hybrid strategy that combines departure timetabling, stop-skipping, and re-routing. In this approach, the single-line bus operational process is modelled as a Markov Decision Process (MDP), where the reward function considers both bus operational costs and passenger waiting time. Using a real-world transportation dataset in Xiamen, China, the experiments verified that our approach is able to reduce the passenger waiting time without higher operational costs and exhibit robustness on bus lines with heavy ridership demand and uneven ridership distribution. This study presents a pioneering endeavour in integrating DRL and transportation geographic information system into bus operation control. The real-time control mechanism enables bus lines to dynamically adapt to ridership demand fluctuations and maintain passenger satisfaction across diverse scenarios.