Multi-Agent Deep Reinforcement Learning (MADRL) for Solving Real-Time Railway Rescheduling Problem

Abstract

The real-time railway rescheduling problem is a challenging task since several factors have to be considered when a train deviates from the initial timetable. Nowadays, the problem is solved by human operators, which is safe but not optimal. This paper proposes a novel state representation for the introduced control problem that enables the efficient utilization of Multi-Agent Deep Reinforcement Learning. To support our claim, a proof of concept network is implemented, and the performance of the trained agent is evaluated. The results show that our approach enables fast convergence and excellent performance, while the representation has the potential for solving much more complex networks.