Solving disruptions in networks with Machine Learning

Among five best in Flatland Challenge by SBB

How can trains learn to coordinate automatically causing minimal delays in large train networks? We choose Reinforcement Learning, a sub-domain of Machine Learning and Artificial Intelligence, to solve this challenge in the “Flatland” competition by SBB. Our trained model dynamically reacts to new situations and solves incidents efficiently with fewer delays.

The Swiss Federal Railways (SBB) operate the densest mixed railway traffic in the world. More than 10,000 trains with 1.2 million passengers and half of Switzerland’s transported goods run on this railway network each day. Due to the growing demand for mobility, SBB needs to increase the transportation capacity of the network. To achieve this, they need to be able to solve unforeseen incidents quickly and efficiently, as not to disrupt larger parts of the network. The vehicle-rescheduling problem arises when a previously assigned trip is disrupted and demands the rescheduling of the trip – including re-routing of possibly other affected trains. The “Flatland” competition aims to address vehicle-scheduling and rescheduling problems, e.g. a traffic accident, a medical emergency, or a breakdown of a vehicle, with a simulator.

Reinforcement Learning to efficiently reschedule a trip

The research group at SBB has developed a platform simulating the dynamics of train traffic and the railway infrastructure. The Netcetera artificial intelligence team at the “Flatland” challenge achieved great results with a model based on a Reinforcement Learning algorithm.

Our AI model had to solve different randomized flatlands in the simulator with unforeseen disruptions. At the more basic level, the model could achieve the goals using ad-hoc decisions. With increasing difficulty from round to round, the model had to be able to plan and react to new and unforeseen circumstances. The automatic algorithm controls the movements of the trains; it is a dynamic model reacting to environmental influences. Our model had to be able to react to disruptions and solve them with a dynamic rescheduling of trains. We solved the challenges successfully and reached the fifth place after the last round at the beginning of January 2020.

Find the technical details about this project in Evgenija's blog post.

Talk to our expert:

Ramon Grunder

Head of Delivery Digital Enterprise

More stories

On this topic