Optimising a Real-Time Scheduler for Indian Railway Lines by Policy Search

The Indian railway network carries the largest number of passengers in the world, with over 8.4 billion transported in 2018, in addition to 1.2 billion tonnes of freight [1]. Nonetheless, the network has only about a tenth the “track-length per passenger” of the U.S., and half that of China [2]. This severe limitation of infrastructure, coupled with variability and heterogeneity in operations, raises significant challenges in scheduling. In this paper, we describe a policy search approach to decide arrival/departure times and track allocations for trains such that the resource and operating constraints of the railway line are satisfied, while the priority-weighted departure delay (PWDD) is minimised. We evaluate our approach on three large railway lines from the Indian network. We observe significant reductions of PWDD over traditional heuristics and a solution based on reinforcement learning.