Learning Railways for Better schedules
Mobility Initiative Project
The problem of organising high-performant railway services is essentially a scheduling problem, assigning scarce infrastructure resources to a series of trains so that both the number of trains (capacity) and performance are maximized. Performance KPIs are considered for both passengers and operators in terms of speed or travel time. Automated scheduling approaches exist in academia and practice. Though, they face a hard trade-off between the speed of computation (the faster, the better); the detail of the mathematical model considered (considering signals); and the scale of the instance solved (one station or one corridor; multiple stations; wide regions or an entire country).
To move forward in this dilemma, we propose to use learning techniques at multiple stages:
- We want to learn the actual KPIs of the traffic controller/planner better. This insight allows us to focus on the actual problems based on the explicit feedback of users.
- We want to use machine learning to provide faster solution approaches. We aim for computational approaches by identifying the most promising research direction while exploring the solutions.
- We want to learn how practitioners accept or modify the solutions computed before implementing them in practice so that we are able to identify and anticipate their implicit preferences.
We plan to integrate this approach in the working environment of both SBB and Siemens to allow direct usability of the resulting algorithms and the applicability of the insights in the current timetabling and scheduling processes.
