Synchromodal freight transport re-planning under service time uncertainty: An online model-assisted reinforcement learning

Yimeng Zhang*, Rudy R. Negenborn, Bilge Atasoy

*Corresponding author for this work

Research output: Contribution to journalArticleScientificpeer-review

28 Downloads (Pure)


The objective of this study is to address the issue of service time uncertainty in synchromodal freight transport, which can cause delays, inefficiencies, and reduced satisfaction for shippers. The proposed solution is an online deep Reinforcement Learning (RL) approach that takes into account the service time uncertainty, assisted by an Adaptive Large Neighborhood Search (ALNS) heuristic that provides state and reward information based on the routing and scheduling. The proposed planning approach re-plans in response to unexpected events and learns from real-time information from various transport modes, including road, railway, and inland waterways. The performance of the proposed planning approach is evaluated in the European Rhine-Alpine corridor under various scenarios with different types and severities of unexpected events. The results demonstrate that the RL approach consistently outperforms other strategies by effectively handling service time uncertainty, leading to reduced costs, emissions, and waiting time, as well as decreased transport delays and improved rewards through accurate decision-making and agile transport re-planning. This study also finds that incorporating event severity information improves the average reward obtained by the RL approach in scenarios involving multiple types of events.

Original languageEnglish
Article number104355
Number of pages33
JournalTransportation Research Part C: Emerging Technologies
Publication statusPublished - 2023


  • Deep reinforcement learning
  • Online transport planning
  • Service time
  • Synchromodal transport
  • Uncertainty


Dive into the research topics of 'Synchromodal freight transport re-planning under service time uncertainty: An online model-assisted reinforcement learning'. Together they form a unique fingerprint.

Cite this