Existing and novel Approaches to the Vehicle Rescheduling Problem (VRSP)

The demands on railway transport are constantly increasing and the transport capacity is reaching its limits. An opportunity to address this challenge is the optimization of traffic planning through novel approaches. In this thesis, existing and novel approaches to two fundamental problems in transport planning are examined with the help of a framework provided by Swiss Federal Railways (SBB).
For the Multi-Agent Path Finding (MAPF) problem, there are different optimal approaches such as Linear Programming, Constraint-based Search, and Operator Decomposition & Independence Detection, as well as suboptimal approaches like Prioritized Planning. The optimal algorithms provide an optimal solution for the problem, but at the expense of a non-polynomial runtime. Instead, a suboptimal approach is to be chosen which provides an appropriate solution within a polynomial runtime.
For the Vehicle Rescheduling Problem (VRSP), various approaches are investigated as well. A special focus is given to the novel Reinforcement Learning approach, where a train should learn independently by means of artificial intelligence how to behave best in a situation. In this thesis, the potential of reinforcement learning is demonstrated, as it can almost keep up with the existing approaches.
An additional challenge is the detection of deadlocks, where several trains inextricably block each other. A complete and reliable detection is again not achievable within a polynomial runtime. Instead, simplified algorithms or again reinforcement learning can be used for a faster but incomplete detection of deadlocks.
In conclusion, it can be stated that reinforcement learning is a novel but promising method, which could be used in railway traffic in an optimizing way. This approach may not be able to replace the existing traffic management methods, but it can at least support these methods.